-
Notifications
You must be signed in to change notification settings - Fork 101
Release Notes
For changes in the Github repository since the last release see here.
In this release we have continued to evolve the support for writing mixed-reality applications using \psi and made further refinements and updates to Platform for Situated Intelligence Studio, the \psi runtime, batch processing infrastructure, as well as other updates and bug fixes across various \psi components.
In addition, we have released SIGMA, standing for Situated Interactive Guidance Management and Assistance, as a testbed application build on \psi that enables research at the intersection of mixed reality and AI. SIGMA is a mixed reality system that leverages the HoloLens 2 device, LLMs, and vision models to guide people through procedural tasks step-by-step. The system is intended to support further research at the intersection of AI and mixed reality. To learn more about SIGMA, please see the corresponding ReadMe.md file. More information about the system is available in this [arxiv paper](add link) and the corresponding extended abstract published at IEEE VR.
Several improvements and extensions have been made to the set of existing mixed reality components:
-
The
PhotoVideoCamera
component can now be configured to turn off continuous auto-focus. Additionally, the camera can be configured in triggered mode (viaVideoStreamUsesTriggeredOutputs
andPreviewStreamUsesTriggeredOutputs
in corresponding configuration). When in triggered mode, the camera will output messages on a cadence specified via the 'TriggeredOutputFrequency' configuration parameter and aligned (in a nearest-message sense) with the first input received on theTriggerInput
receiver. -
Added the option in
PhotoVideoCamera
to output either NV-12 encoded images or BGRA images (thanks @austinbhale). -
When exporting HoloLens data, camera intrinsics are now exported for every frame, since they may periodically change due to continuous auto-focus.
-
The
Microphone
component now supports configurableMediaCategory
andAudioProcessing
modes. Note that if not specified, theMediaCategory
defaults toMediaCategory.Speech
whereas previously it defaulted toMediaCategory.Other
. -
The
WinRT.GazeSensor
component now outputs a stream of tuples of eye gaze and head pose. -
The
SpatialSound
component now has an output stream representing the audio buffers that were played through the speakers, with originating times corresponding to actual playback time. -
The
Handle.cs
renderer component uses the backend OpenXR time when posting its pose.A few new components have also been added: -
Added a generic
ModelRenderer
component for rendering any StereoKitModel
. -
Added a new StereoKit rendering component,
OpenXRHandRenderer
, which renders OpenXRHand
objects from any input data stream.
Several other infrastructural updates include:
-
The approach for setting and handling world transforms and spatial anchors has been refactored and improved. Local as well as persistent Azure Spatial anchors are supported via the
LocalSpatialAnchorProvider
(previously namedSpatialAnchorHelper
- thanks @austinbhale for past contributions to this) andAzureSpatialAnchorProvider
classes respectively. Both providers can be used to set the world coordinate system via theMixedReality.SetWorldCoordinateSystem
method. -
A static method for loading 3D models stored in a HoloLens folder as a StereoKit
Model
is now available. -
Serialization of
AudioBuffer
inHoloLensCaptureInterop
now includes the entireWaveFormat
header instead of assuming a fixed format. -
Operators in
HoloLensCaptureExporter
have been moved toHoloLensCaptureInterop
.
Several updates have been made to the set of existing visualizers, and new infrastructure has been added to support visualizers that render a temporal segment of stream data. A detailed list of these changes is below:
-
Updated 2D and 3D visualization infrastructure with support for stream interval visualizers. This is based on a newly introduced
XYIntervalVisualizationObject
andXYZIntervalVisualizationObject
types (deriving from StreamIntervalVisualizationObject) that allows the user to specify which portion of the stream is used to construct the visualization data (currently
All,
Selection,
View, and
CursorEpsilonare implemented). As an example, the
LabeledPointIntervalVisualizationObjectand
Point3DIntervalAsPointCloudVisualizationObject` allows for visualizing a scatter plot of points from a stream with 2D or 3D point messages. These visualizers provide a basis for trajectory visualizations. -
Added new 2D visualizers for
Line2D
and a corresponding stream interval visualizer. -
Added new 2D visualizer for
Object2DDetectionResults
-
Changed
PseudoColorize
method used by depth image visualizer to not render pixels outside the specified range. -
Expanded the set of diagnostic visualization options for streams to include message dropped and processed percentages, both total and withing the averaging window.
Several other improvements have been made to PsiStudio and the visualization infrastructure:
-
Added Help window in PsiStudio (accessible from toolbar button) that displays a list of useful mouse and keyboard shortcuts.
-
Separated out the most recently used lists for datasets and stores.
-
Added a setting to prompt/auto-save changes made to the current layout.
-
Added support for copying time interval annotation boundaries from other visualized annotation streams.
-
Added support for deleting partitions from all sessions.
-
Added support for navigating to a specified session and time by using a
<session_name>@<time>
syntax in the Go-To-Time dialog box. -
Added several key shortcuts:
Space
shortcut for play/pause,Shift
+Left/Right Arrow
shortcut for moving the selection to the next annotation. -
Added menu item for "Open Session Folder in Explorer" on sessions and "Open Partition Folder in Explorer" on partitions.
-
Settings file is now in a machine-specific folder, under
Documents\PsiStudio\Settings\{machine_name}
. -
Simplified running of UI actions within a progress window with new static methods in
ProgressWindow
. -
Fixed an issue with the alignment of audio playback in PsiStudio: now PsiStudio correctly plays back the audio based on originating times.
-
Fixed audio channel selection in PsiStudio and added a setting to control whether to playback only the selected channel.
Several fixes and improvements have been made to the \psi runtime:
-
Added method
OpenStreamOrDefault<T>
forImporter
andSessionImporter
. The method opens the specified stream if the stream exists or returns null otherwise. -
Added
DeliveryPolicySpec
enum type to enable specifying pipeline-level delivery policies during batch task processing. -
Added support for optional progress reporting in synchronous
Pipeline.Run()
methods. -
Improved the memory allocation strategy in
BufferWriter
. -
Fixed an issue with the importer initializing
KnownSerializers
based on the evolving set of default serializers. Moving forward, the runtime still allows for constructing aKnownSerializers
instance that uses an existing set of known serializers as a starting point (by passing them as a second parameter in the constructor) -
Fixed an issue where custom serializers based on
BackCompatSerializer
sometimes failed to deserialize data from stores. -
Fixed an issue where
IndexEntry
metadata was incorrectly interpreted when reading from stores created with different .NET versions. -
Fixed an issue in
InfiniteFileWriter
where the file handle is not disposed when out of disk space (thanks @austinbhale). -
Fixed an issue enumerating stores with names containing whitespace (thanks @austinbhale).
-
Fixed an issue in pipeline diagnostics capture for dropped messages that in some cases led to incorrect statistics being displayed.
-
Added
Interval<>.Intersection()
method for computing the intersection of a set of intervals. -
ConfigurationHelper
is now a static class with new static methodsReadFromFileOrDefault<T>
andReadFromStream<T>
to read a configuration from a file or stream, respectively. -
Introduced mechanism for accessing platform resources (such as image encoders/decoders) via the
PlatformResources.Get<>
method in the runtime. Resources have to be registered viaPlatformResources.Register<>()
. -
An optional
maxSpan
parameter was added to theAdjacentValuesInterpolator
allowing users to control the maximum timespan between adjacent values for which the interpolator will produce a result.
Several improvements and extensions have been made to the batch task processing infrastructure:
-
The method-based approach to defining batch tasks has been deprecated, and moving forward batch tasks will be defined via the classes derived from
BatchProcessingTask<TBatchProcessingTaskConfiguration>
. The batch task infrastructure was extended with handlers that are called prior to and after running on each session and dataset, and on encountering exceptions or cancellation, to enable initialization and clean-up code. Batch tasks can now be run programmatically onDataset
andSession
objects. -
Added
StreamCollection
class to support a pattern for hierarchical compositing of batch processing tasks. -
Removed the need for declaring the
[ExpandableObject]
attribute on batch task configuration classes -
Batch tasks can now be more easily implemented as part of .NET standard libraries and they can leverage platform specific resources via the
PlatformResources
class methods. -
An overloaded version of the
ToTimeIntervalAnnotation()
stream operator has been added to support conversions from a 3oolean stream into a time interval annotation stream. -
Several methods on
Dataset
andSession
have been renamed and/or made async.
- Added a method for computing an interpolated pixel in a depth image. This is useful when intersecting a 3D ray with a depth map that might have holes (pixel values of 0).
-
Extended
Mesh3D
to include an (optional) overall pose, and added algorithms for determining intersection withRay3D
and for determining the closest point on the mesh to a specifiedPoint3D
-
Static methods for computing triangle planes and testing whether a triangle encloses a point.
-
Added a method on
Box3D
for computing a closest point. -
Added operator for interpolating between two
Ray3D
s and determining whether a triangle contains a point.
- Added a new
Microsoft.Psi.Media
library defining the platform-independent interfaceImpeg4Writer
.
-
Added a constructor that creates an empty
AudioBuffer
with a specifiedTimeSpan
and format. -
Added a concatenation (
+
) operator forAudioBuffer
. -
Added a
SelectChannel
operator forAudioBuffer
.
-
Added
SpeechSynthesizer
component providing access to Azure text-to-speech services. -
Removed obsolete
BingSpeechRecognizer
component which has been replaced byAzureSpeechRecognizer
. -
Added an operator for converting a stream of
SpeechSynthesisProgress
messages to temporal annotations.
- Due to deprecation of the method-based approach to defining batch tasks, the method-based
PersonGroupTasks
have been removed.
- Added a new
IInteropSerializable
interface, which, together with a new set of operators for reading and writingIInteropSerializable
types from/to binary writers enables a simpler pattern for interop serialization.
-
Added \psi components for running the Detic and SEEM models. The components connect over a zero-mq connection to a python server that runs these models. The server functionality is provided at under Models in the Integrations folder. For installing / configuring these python servers, see the documentation available in the corresponding readme files for Detic and SEEM
-
The
OnnxModel
base class now takes generic input and output type parameters to support models with input and output data types other thanfloat
.
-
Updated code analysis in projects to use .NET analyzers (version 8.0.0).
-
Upgraded test projects to target .NET 8 and updated them to use the latest MS Test NuGet packages.
-
Upgraded C++ projects to target the latest Windows SDK version (currently 10.0.22621.0)
-
Upgraded UWP projects to target Windows SDK version 10.0.22621.0.
-
Removed obsolete
MicrosoftSpeech
project and associated components.
In this release we have continued to evolve the support for writing mixed-reality applications using \psi and included several important updates to Platform for Situated Intelligence Studio and the \psi runtime, as well as other updates and bug fixes across various \psi components.
We have continued to evolve support for writing mixed-reality applications with \psi. This release includes new components for more direct and fine-grained access to hand tracking information via OpenXR APIs and gaze tracking information via WinRT. The update also includes a refactoring of various classes and namespaces under Microsoft.Psi.MixedReality
for more clarity. A more detailed list of changes is below:
- Added a new WinRT-based source component for capturing eye and head gaze:
GazeSensor
in theMicrosoft.Psi.MixedReality.WinRT
namespace. - Added a new OpenXR-based
HandsSensor
component in theMicrosoft.Psi.MixedReality.OpenXR
namespace. - The HoloLensCaptureApp now uses the WinRT-based
GazeSensor
and the OpenXR-basedHandsSensor
components. -
BREAKING CHANGE: Refactoring of WinRT/OpenXR/StereoKit/MediaCapture sensors and data types:
- Reorganized
Microsoft.Psi.MixedReality
sensor components and datatypes into different namespaces reflecting the underlying technology used:Microsoft.Psi.MixedReality.StereoKit
Microsoft.Psi.MixedReality.OpenXR
Microsoft.Psi.MixedReality.WinRT
Microsoft.Psi.MixedReality.ResearchMode
Microsoft.Psi.MixedReality.MediaCapture
- The following classes/types from the
Microsoft.Psi.MixedReality
namespace have been moved to a different namespace and/or renamed:-
Hand
,EyesSensor
,Handle
,SpatialSound
,Microphone
have been moved to theMicrosoft.Psi.MixedReality.StereoKit
namespace. - All StereoKit renderer components have been moved to the
Microsoft.Psi.MixedReality.StereoKit
namespace and renamed from <Name>StereoKitRenderer to <Name>Renderer, e.g.,Box3DStereoKitRenderer
->Box3DRenderer
. -
DepthCamera
,Accelerometer
,Gyroscope
,Magnetometer
,VisibleLightCamera
have been moved toMicrosoft.Psi.MixedReality.Research
namespace. -
MediaCaptureMicrophone
has been renamed toMicrophone
and moved under theMicrosoft.Psi.MixedReality.MediaCapture
namespace. -
PhotoVideoCamera
has been moved to theMicrosoft.Psi.MixedReality.MediaCapture
namespace.
-
- Reorganized
-
BREAKING CHANGE:
MixedReality.InitializeAsync()
is nowMixedReality.Initialize()
(no longer async). In general, make sure that all UWP apps deployed to the HoloLens use avoid Main()
method, and notasync Task Main()
. - Updates to
HandVisualizationObject
to no longer require adapters. - Created a batch processing task for projecting 3D hand poses into the 2D image space of available camera views, for visualization in PsiStudio:
ProjectHandsToCamerasTask
. - Further expanded the set of format serializers in
HololensCaptureInterop
Some of the primary changes to PsiStudio involve additional support for expanding streams (e.g., key expansion for Dictionary streams and script-based expansion for computing derived streams), the ability to save batch task configurations, and a new settings dialog. A number of other improvements and fixes have also been made. A more detailed list of changes is below:
- Added key expansion for Dictionary streams.
- Added script-derived streams.
- Batch processing task configuration parameters may now be saved and loaded.
- Added setting that enables skipping the security warnings upon loading 3rd party code.
- Added support for editing the settings via a dialog from within PsiStudio.
- Added setting that enables skipping the security warnings upon loading 3rd party code.
- Added support for remembering a list of most recently used dataset filenames.
- Added visualizer for dictionary of point clouds.
- Removed
AnimatedModel3DVisualizationObject
and replaced with a newPosedModelFromFileVisualizationObject
. This new visualizer works over streams ofCoordinateSystem
poses, and will render a 3D model at the given pose according to a property for configuring which model file to load from disk (.obj, .stl, .3ds, .lwo, .objz, .off, and .ply files are supported). - Added keyboard shortcuts for Zoom to Selection (Ctrl+S) and Zoom to Session (Ctrl+A)
- Added support for removing a specified partition from all sessions.
- Added support for copying the cursor time as ticks to clipboard.
- Added more copy-to-clipboard commands for partitions, sessions and streams in the dataset tree view.
- Changed the settings file to save AdditionalAssemblies as a collection instead of a string.
- Added support for type mappings in PsiStudio settings, which allows for loading streams of an unknown type under a known type. This feature is particularly useful when datatype names have changed, as PsiStudio needs to know this information when loading the store.
- Fixed a bug in expanding members for types that contain members that are nullable.
- Fixed an issue with refresh in pipeline diagnostics visualization.
- Additional fixes and cleanup in the computation of various source stream and subsumed stream properties on the stream tree nodes.
- Refactored the
StreamTreeNode
infrastructure to support more general derived streams. Stream bindings (which are persisted in layouts) can now reflect a binding to a derived stream. TheStreamBinding
stores both a derived stream adapter and a visualizer stream adapter and these two adapters are chained together (via theChainedStreamAdapter
class) to effect the binding to the stream. - Refactored the construction of context menus for visualizers in PsiStudio by moving it from the views to the visualization objects. This puts it more in line with an MVVM approach and unifies the previous separate approaches we had for 3D and 2D visualizers.
- Moved (and refactored) some of the helper operators for accessing collections of time interval annotation from the annotation visualizer into Microsoft.Psi.Data\Annotations and Helpers.
- The
ImagePool
class now allocates different subpools based on binning the allocation size, leading to improved performance.
- BREAKING CHANGE:
CalibrationExtensions.IntersectLineWithDepthMesh()
is nowCalibrationExtensions.ComputeRayIntersection()
. The functionality is effectively the same, (intersecting a ray, rather than a line, with a depth mesh), but is more robust for different kinds of depth images. -
CameraIntrinsics.GetPixelPosition()
now returns null if the 3d point passed in is behind the camera.
- Added Volume property on
Bounds3D
andBox3D
. - Added support for exporting point clouds in PLY format from the
DepthImageCameraViewAsPointCloudVisualizationObject
.
- Addressed concurrency issue in
PsiStoreWriter
that occured when multipleWrite()
commands would be generated while the pipeline is running, fromParallel()
operators in the pipeline. - Addressed concurrency issue occurring when estimating pipeline progress when new receivers are dynamically added to a component while the pipeline is running.
- Enabled the use of
AdjacentValueInterpolator{T}
inside Fuse/Join operators. - The
ThrottleWhenFull
boolean property (and corresponding construction parameter) ofDeliveryPolicy
has been changed to anint?
property namedThrottleQueueSize
. This allows more flexibility to specify the queue size threshold at which throttling occurs. - The behavior of delivery policies with throttling has changed. Previously, these policies would cause incoming messages to be dropped at the receiver if the upstream source attempts to post messages while throttled. These messages are no longer dropped but instead will be queued at the receiver.
-
BREAKING CHANGE: The
autoClone
flag has been removed from thePipeline.CreateReceiver
methods. Receivers will always receive a cloned message which guarantees that any changes made to the message contents in the receiver are isolated from the sender's copy of the message. Cloned messages will be automatically recycled when the receiver method returns. If the message contents need to be retained beyond the lifetime of the receiver method, an explicit copy should be made using theDeepClone
extension method. - Added support for caching in
TypeResolutionHelper
class for performance improvements. - Added support for type name synonyms with
KnownSerializers.RegisterDynamicTypeSchemaNameSynonym(...)
. These are used when parsing type schema information toPsiStore.OpenDynamicStream(...)
. - Added new
BackCompatClassSerializer
andBackCompatStructSerializer
classes to simplify writing a custom serializer that needs to maintain backward compatibility with previous versions of a data type serialized with the auto-generated default serializer. For an example, see theAngularVelocity3D
class in Microsoft.Psi.Spatial.Euclidean.
- Fixed optional FFmpeg support on Linux and removed obsolete
Microsoft.Psi.Media.Native.x64
project. - Modified the Windows
Mpeg4Writer
component to buffer and write audio and video in a coordinated way.
- Fixed an issue re: disposing the underlying resources in various model runner components.
-
BREAKING CHANGE:
MaskRCNNModelRunner
now internally resizes input images to match the configured size and resizes outputMaskRCNNDetection
results to match the original input image size. -
MaskRCNNModelConfiguration
now includes aConfidenceThreshold
used in filtering results. - Created templated version of the
OnnxModelRunner
class, enabling users to provide input and output adapters via the constructor.
- Added support for setting annotation boundaries to selection boundaries via context-menu.
- Improved performance for multi-track annotation visualizer.
- Multi-track annotation visualizer support now showing different subsets of tracks via the
Show Tracks
property. Users can select whether to show all tracks, only tracks with events in view, or manually specify (via theShow Tracks Selection
property) which tracks are to be shown.
This release includes important updates to mixed reality support in \psi, including a set of tools for streaming data from a HoloLens 2 to a separate PC for data collection and export. The release also includes several updates to visualization and PsiStudio, the addition of a wrapper for running MaskRCNN models, updates to Azure Kinect Components, as well as some runtime updates and various other bug fixes.
- Introduced
ResearchCameraConfiguration
base type; nowVisibleLightCameraConfiguration
andDepthCameraConfiguration
derive from this base type; in the process renamedMode
configuration property toDepthSensorType
andVisibleLightSensorType
respectively. - Added
GetCurrentTimeFromOpenXr()
extension method on pipeline for retrieving the current time from OpenXR. - Updated originating time computation at various mixed reality sensor components to more accurately reflect capture time.
- Added capability (and corresponding configuration option) for outputting camera view streams (using the new
*ImageCameraView
types) from the photovideo, depth, and visible light cameras. - Changed
PhotoVideoCamera
to emit encoded NV12 images or image camera views (for increased performance). -
DepthCamera
andVisibleLightCamera
now allow for calling aCalibrate()
method prior to starting the pipeline. The calibration process is compute intensive and sometimes it is desirable to compute the calibration information prior to starting the pipeline to avoid slowdowns on pipeline startup. - Refactored and simplified the set of operators available for converting poses to/from StereoKit.
- Added
PsiInput
static class that provides head, hand, and eye inputs in \psi world coordinates. - Established a pattern for creating new StereoKit renderers (deriving from
StereoKitRenderer
), and added a few more. - Changed
EyesSensor
output representation from aCoordinateSystem
to aRay3D
. - Added an
Out
emitter to theHandsSensor
that carries both hands in a tuple (Left, Right).
- Added a set of tools that enable capturing sensor data from the HoloLens 2 device into \psi stores, and for exporting these stores to other formats.
- HoloLensCaptureApp: is an application runs on the HoloLens, capturing and remoting data streams.
- HoloLensCaptureServer: is an application that runs on a Windows machine, receiving remoted data streams from the HoloLensCaptureApp and writing them to a \psi store.
- HoloLensCaptureExporter: is a tool that can convert the data captured in the \psi stores to other formats.
- Updated Nuget references
Microsoft.Azure.Kinect.Sensor
to 1.4.1 andMicrosoft.Azure.Kinect.BodyTracking
to 1.1.1, which bring several improvements in the underlying SDKs, including no longer requiring CUDA dependencies and supporting more GPU setups. - Added an additional configuration option to the
AzureKinectBodyTracker
component for optionally using the "lite" version of the underlying pose estimation model. According to the Azure Kinect documentation, "this model trades ~2x performance increase for ~5% accuracy decrease."
- Added
Description
member onIImageToStreamEncoder
andIDepthImageToStreamEncoder
; this in turns allows more informative displays in diagnostics visualization when using encoders. - Added
IImage
andIDepthImage
interfaces. - Renamed
EncodedImage.SetBuffer(...)
, which copys from abyte[]
, toEncodedImage.CopyFrom(...)
. - Added
EncodedImage.CopyFrom(...)
overload allowing data to be copied from a memory pointer. - Separated JPEG/PNG decoders into
ImageFromBitmapStreamDecoder
in each ofMicrosoft.Psi.Imaging.Windows
andMicrosoft.Psi.Imaging.Linux
- Separated GZip decoder into
ImageFromGZipStreamDecoder
in cross-platformMicrosoft.Psi.Imaging
- Added NV12 decoder
ImageFromNV12StreamDecoder
in cross-platformMicrosoft.Psi.Imaging
-
ImageFromStreamDecoder
now supports NV12 encoding. - Renamed type
DepthPixelSemantics
toDepthValueSemantics
; the depth pixel semantics information has been migrated in a breaking change from theCameraIntrinsics
type to theDepthImage
type, which now stores two new members:DepthValueSemantics
andDepthValueToMetersScaleFactor
. - Fixed a bug in
CalibrationExtensions.CalibrateCameraIntrinsicsAndExtrinsics()
where the bottom-right element of the resulting intrinsics matrix was not properly being set to a value of 1.
The release includes several usability updates for PsiStudio, as well as a few adapters, visualizers and infrastructural updates:
- Added support for moving the selection to the right or the left (via Alt+Left/Alt+Right, or toolbar buttons).
- Added support for switching the currently visualized session by double-clicking on it.
- Added "Go To Time" button on toolbar, also accessible via Ctrl+G shortcut and from the context menu on any panel.
- Double-click on a regular stream in the datasets menu expands the stream members.
- Added the ability to set the default cursor epsilon on instant visualization panels.
- Added property on labeled rectangle visualizers to show/hide the label.
- Sessions now support adding multiple partitions from a specified folder.
- Dataset now supports adding session (with multiple partitions) from a specified folder.
- Dataset now supports adding multiple sessions (with multiple partitions) from a specified folder.
- Expanded support to copy to clipboard to include session name and/or cursor time.
- Added support for switching which session is being visualized from the context menu of a partition or stream in a non-visualized partition.
- Added adapters to visualize
Mesh3D
objects as point clouds (showing only the vertices). - Added support for visualizing lists of
Ray3D
objects. - Added support for naming stream adapters via the
StreamAdapter
attribute and constructing better stream context menu item names when multiple visualizers with different adapters with the same type signature are available. - Addressed an issue that was creating a 2x slowdown in rendering visualizations that use a canvas view.
- Fixed a bug when copying cursor time to clipboard.
- A few small enhancements to text visualizers.
- Converted
Box3D
type from struct to class and updated corresponding visualizers. - Added
Scale()
method onBox3D
andBounds3D
. - Added a public constructor for
PointCloud3D
from matrix. - Added epsilon tolerance parameters to
AngularVelocity3D
,CoordinateSystemVelocity3D
constructors andBounds3D.ContainsPoint()
method. - Added
Mesh3D.ToPointCloud3D()
method to convert mesh vertices to a point cloud. - Added
ImageCameraView
types containingShared<Image>
along with camera intrinsics (ICameraIntrinsics
) and pose (CoordinateSystem
).- Added corresponding
EncodedImageCameraView
,DepthImageCameraView
, andEncodedDepthImageCameraView
and PsiStudio visualizers. - Added adapters from
(<image>, ICameraIntrinsics, CoordinateSystem)
to new*ImageCameraView
types.
- Added corresponding
- Added support for naming components and stream operators, as well as default names for most components and stream operators. This enables a more readable display of component names when visualizing pipeline diagnostics.
- Stores containing streams of
Dictionary<TKey, TValue>
are now cross-platform compatible. - Fixed an issue in the store writer
OpenStream()
method regarding setting the opened time (which ensures correct behavior in batch processing / replay). - Fixed an issue where pipeline shutdown was sometimes being delayed due to a pending message with an originating time in the future.
- Renamed
AnnotationSchema.Load()
toAnnotationSchema.LoadFrom()
and addedTryLoadFrom()
method. - Refined/updated overloads for
ToTimeIntervalAnnotations()
stream operator (used for constructing time interval annotations from various source streams).
- Added stream operators for exposing various image analysis results.
-
TcpSource
now supports custom, user-specified deallocators to manage the lifetimes of the objects it creates.
- Added support for using shape dictionaries when running onnx models.
- Added Mask R-CNN model runner (object detection, classification, bounding boxes and masks) and a PsiStudio visualizer.
- Updated pipeline visualization to show the type of interpolator used by
Join
andFuse
as a tooltip on hover over the respective component. - Added statistic for percentages of messages dropped to pipeline diagnostics visualization.
This release includes additional \psi libraries and components for writing \psi mixed-reality applications for HoloLens 2, new types and operators for representing and dealing with various spatial geometric objects, numerous updates and fixes to Platform for Situated Intelligence Studio (from updates to user interface to new and updated visualizers and support for multi-track temporal annotations), a few updates to some of the stream operators in the \psi runtime, as well as a number of other component fixes and updates. The changes are described in more detail below.
Platform for Situated Intelligence now supports the development of mixed reality applications. This release includes several components for accessing and processing the rich multimodal sensor streams (such as head tracking, hand tracking, gaze tracking, RGB and depth cameras, IMU, etc.) as well as example \psi components for rendering output on the device. For now, only the HoloLens 2 device has been tested and is supported, but in theory, it should be possible (with perhaps a little extra effort) to target other mixed reality platforms such as Oculus Quest, Windows Mixed Reality, Oculus Desktop, SteamVR, etc. Integration between \psi and the HoloLens 2 is enabled by two primary underlying technologies: Research Mode and StereoKit.
This release includes a new project and namespace: Microsoft.Psi.Spatial.Euclidean
, which extends the set of types available in MathNet.Spatial.Euclidean. New classes and operators allow for representing, manipulating, and visualizing streams of 3D rectangles, boxes, point clouds, meshes, velocities, and more.
The updates to visualization and PsiStudio include support for multi-track annotations, improvements to the user interface, new visualizers and adapters, updates to existing visualizers, some updates and code refactoring in the visualization infrastructure, as well as bug fixes and performance improvements.
-
Support for multi-track annotations:
- Added support for manually constructing overlapping time-segment annotations in a single annotation stream. Context menus have changed on annotation stream visualizations to support creation of annotations on multiple tracks. This is a BREAKING CHANGE with respect to the definition of annotation schemas and the data types involved in persisting annotations.
-
Updates and improvements in the user interface:
- Added support for Dataset auto-save in PsiStudio -- see this pull request.
- Added support in PsiStudio to automatically open the last loaded dataset on startup. This is configured with a settings option. Users can now toggle both this option and the option to autosave datasets from the PsiStudio File menu.
- Added shortcut buttons on panels for showing/hiding them.
- 2D visualizers can now be zoomed into using the mouse wheel. Users can pan the visualizer by dragging the mouse with the right mouse button down.
- Users can now set a Threshold Value on Timeline Panels such that any values either above or below this threshold is highlighted to make it easier to find areas in the data that may require further scrutiny.
- Added support for displaying throughput statistics (per hour, min, sec.) in dataset / session / partition / streams tree.
- Added new properties on 3D visualization panels. These include whether to zoom/rotate around the mouse point (which is now the default mode), the zoom and rotation sensitivity, as well as the options to display various info for the viewport, such as FPS, number of triangles, coordinate system, view cube, etc.
- Added support for deleting layouts from the layout toolbar.
- When attempting to save a layout under a name that already exists, an option to overwrite or cancel is now presented (rather than just an error message).
- Updated the mechanism by which PsiStudio detects live stores.
- Added ability to set the cursor epsilon on instant visualizers directly from the visualizer context menu.
- Added visualizer context menu option to copy the cursor time to clipboard.
-
New visualizers and adapters:
- Added a new 3D visualizer for displaying text in a 3D location.
- Added an instant visualizer for
IFormattable
objects that allows to specify the format string as a property and renders the object as text. - Added
MathNetPoint2D
andNullablePoint2D
to scatter plot stream adapter for visualization. - Added support for XY panels and Canvas panels. The former can include visualization objects that display (x,y) data, such as the scatter plot visualizer, whereas the second can include visualization objects that display any data on a flat canvas without axes, such as the text visualizer.
- Added better support for visualizing booleans and series of booleans, i.e.,
Dictionary<Key, bool>
, via plot visualizer.
-
Updates to existing visualizers:
- Added support for font size and border thickness in billboard visualizer.
- Added support for XY charts in
Microsoft.Psi.LiveCharts.Visualization
. - Added support for visualizing 3D graphs based on a specified node visual 3D and edge visual 3D via
ModelVisual3DGraphVisualizationObject
. - Added support for displaying lossy delivery policies as dotted in the pipeline diagnostics visualizer (enabled by default).
- Changed delivery policy highlight in pipeline visualizer to use transparency rather than color.
- Updated edge labels in pipeline visualizer to use the heatmap color when using heatmap statistics. Updated default heatmap statistics to "none".
- Added scroll bar capabilities for the text and
IFormattable
visualizers. - Added support for visualizing images in
RGB_24bpp
format -- see this pull request. - Added support for displaying (as a readonly property) the value of the pixel under mouse in the depth image visualizer.
- Added support for controlling the range mode via context menu on the depth image visualizer.
-
Updates to visualization infrastructure:
- Refactored 2D visualization panels into two types:
XYVisualizationPanel
(which supports visualizers that present XY data) andCanvasVisualizationPanel
(which supports 2D visualizers that do not present XY data). - Systematized and refactored base classes for collections of 3D visualization objects.
- Systematized and expanded support for batch processing tasks to allow for task parameterization and parameter validation.
- Added support for creating visualizers for interface types.
- Enabled visualizers derived from
StreamIntervalVisualizationObject
to specify which data interval should be read in (by overriding a virtualGetTimeInterval()
method.)
- Refactored 2D visualization panels into two types:
-
Bug fixes and performance improvements:
- Fixed a bug that crashed CartesianChartVisualizationObject.
- Improved performance on point cloud visualizer by using the camera space mapping matrix.
- Small fix in diagnostics viz to display
Fuse
components with a "+", likeJoin
. - Changed and fixed some problems in approach for reordering panels in the visualization container. To reorder panels, left-click the handle that appears at the left edge of the panel and drag up or down. Left-click and drag elsewhere in timeline panels now only moves the timeline.
- Added
Last()
andLast(count)
stream operators, as well as support onProcessor
component for an optional action to execute when the input stream closes. - Refactored the
Interpolator<TIn,TOut>
class hierarchy to allow for implicit operator conversions fromRelativeTimeInterval
andTimeSpan
toInterpolator<T>
. This allows for these conversions to be used inFuse()
operator calls, not justJoin()
. - Systematized the vector join operations (in the process removed the sparse vector join operator).
- Added
RemoteClockExporter
/RemoteClockImporter
components to allow for synchronizing a local pipeline clock with a pipeline running on a remote machine (by exchanging clock information and applying a local offset). - Added
IsClearRequired
toISerializer<T>
interface. - Updates to
PipelineDiagnostics
data structure; these updates constitute a BREAKING CHANGE for serialization of pipeline diagnostics.
-
BREAKING CHANGE: Renamed
ToPixelFormat()
stream operator toConvert()
. -
BREAKING CHANGE: Removed
ToGray()
stream operator - use theConvert()
operator instead. - Added
FromFile()
andSave()
methods toImage
class. - Added
Convolve()
operator (for now only for grayscale images) - Added support for clipping the crop region to image boundaries during
Crop()
operations. - Added
SetBuffer()
methods to theEncodedImage
andEncodedDepthImage
classes. - Added additional image drawing methods and operators, including
DrawText()
,FillRectangle()
, andFillCircle()
. - Additional image drawing operators -- see this pull request.
- Depth images now support encoding/decoding via the TIFF compression format -- see this pull request.
- Added
ImageToGZipStreamEncoder
component andEncodeGZip()
stream operator to support GZip compression forShared<Image>
- Added a
WaveStreamSampleSource
component that can produce on-demand an audio sample specified by aSystem.IO.Stream
.
- Implemented
IEquatable
interface onCameraIntrinsics
and addedGetCameraSpaceMapping
method that returns a mapping matrix that can be used to quickly transform depth value pixels into 3D space. - Renamed some of the members of
ICameraCalibration
:ToPixelSpace()
->GetPixelPosition()
,ToCameraSpace()
->GetCameraSpacePosition()
,GetCameraSpaceMapping()
->GetPixelToCameraSpaceMapping()
,DistortPoint()
->TryDistortPoint()
,UndistortPoint()
->TryUndistortPoint()
-
CameraIntrinsics.GetPixelPosition()
now returns aPoint2D?
which isnull
if the pixel is outside the field of view of the camera (unless optional argumentnullIfOutsideFieldOfView=false
). -
DepthDeviceCalibrationInfo.ToColorSpace()
renamed toGetPixelPosition()
and now returns a nullablePoint2D?
which isnull
if the pixel is outside the field of view of the camera (unless optional argumentnullIfOutsideFieldOfView=false
). - Added
CameraIntrinsics.TryGetPixelPosition()
andDepthDeviceCalibrationInfo.TryGetPixelPosition()
. - Added a
DepthPixelSemantics
field which indicates whether the depth values represent distance from camera to point, or distance from camera to the camera-axis-perpendicular plane containing the point. - Added methods in
CalibrationExtensions.cs
for calibrating a camera from sets of 2D image points and corresponding 3D points (intrinsics and/or extrinsics). PreviousComputeCameraIntrinscs()
methods are now calledCreateCameraIntrinsics()
.
- Added support for defining configurable batch processing tasks via classes that derive from
BatchProcessingTask<TBatchProcessingTaskConfiguration>
. - Renamed
SessionImporter.HasStream()
toSessionImporter.Contains()
for consistency with other APIs. - Removed
UseRelativePaths
property fromDataset
class -- see this pull request.
- Added
RemoteClockImporter
andRemoteClockExporter
components to synchronize pipeline clocks between \psi pipelines running on different machines. - Added
TcpWriter
andTcpSource
components toMicrosoft.Psi.Interop
allowing for primitive transport over TCP using existing format providers (e.g. JSON, MessagePack) or using custom implementations ofIFormatSerialize
/IFormatDeserialize
- Added Rendezvous system by which \psi apps may publish and discover available
NetMQWriter
/TcpWriter
/RemoteExporter
/RemoteClockExporter
endpoints and available streams.
- In
ImageNetModelRunnerOutputParser
class, removed theGetPredictions()
method and added instead aGetLabeledPredictions()
method that returns all labeled predictions, and aGetTopNLabeledPredictions()
method that returns the top-n labeled predictions.
A new project and namespace: Microsoft.Psi.Filters
, contains some simple initial filters that operate over streams of doubles. It currently contains the following filters: Low-pass (MathNet), High-pass (MathNet), Band-pass (MathNet), Band-stop (MathNet), Denoise (MathNet), OneEuro.
- The
Microsoft.Psi.MicrosoftSpeech.Windows
library has been deprecated as of this release. Existing users should migrate to theSystemSpeechRecognizer
component available in theMicrosoft.Psi.Speech.Windows
project.
This release includes various updates to the runtime, tools, and components. Numerous improvements have been made in Platform for Situated Intelligence Studio, including new visualizers and improvements to existing visualizers, new features and improvements to existing features that accelerate working with the tool, improvements in the underlying data access and visualization infrastructure, and other bug fixes. In addition to the changes to PsiStudio, several updates and bug fixes have also been made to the runtime, to diagnostic metrics capture and visualization, and to imaging and calibration components and APIs.
This release includes numerous updates to Platform for Situated Intelligence Studio, falling into 4 four broad categories: (1) new and updated visualizers, (2) new features for improved usability, (3) changes to infrastructure, and (4) other bug fixes.
(1) New visualizers and updates to existing visualizers:
- Added a visualizer for numeric series, i.e., streams containing
Dictionary<TKey, TNumeric>
, whereTNumeric
is any numeric type (e.g., int, long, double, bool, decimal, and corresponding nullables.) The dictionary keys are automatically converted to strings and act as a series name. A plot is generated for each series. - Added a visualizer for a list of
List<Point3D>
as a point cloud. - Added a visualizer for depth images which shows them as pseudo-colorized images.
- Added new visualizers and adapters for spatial camera views.
- Added a visualizer that displays strings as text in a 2D canvas.
- Added a visualizer that displays a string with position in a 3D billboard.
- Added support for displaying image resolution and mouse position in the properties of image and depth image visualizers.
- Fixed issues with tooltip rendering on nodes in diagnostics visualization.
- Added a visualizer for MathNet Ray3D objects.
(2) New features and improvements to existing features:
- X-Axis and Y-Axis scale properties have been moved from each individual timeline and 2D visualization object to the corresponding visualization panel. As a result, all visualization Objects in timeline or XY panels now automatically use the same horizontal and vertical scale that has been set via properties of the panel.
- Added support for displaying additional dataset, session, partition, and stream information in the Dataset panel, to the right of the tree nodes.
- Added hover-buttons for show/hide and remove visualization objects in the layouts tree, and for closing visualization panels.
- Added "Expand All Nodes" and "Collapse All Nodes" options on stream tree node context menu.
- Added a toolbar button in the Layout tab for clearing the layout.
- Updated playback speed textbox on the toolbar to support speed factors larger than 11. The corresponding increase/decrease buttons double or half the speed factor.
- Redesigned the context menu system in the Visualizations tab. Visualization panels and visualizers now all contribute to the context menu in a structured way.
- Added support for visualizing diagnostics information as streams, by right-clicking on an edge in the diagnostics visualizer and creating specific derived stream nodes.
- Added support for allowing mouse to move without the cursor following the mouse if Shift key is held down.
- Updated default behavior on opening a dataset to expand sessions only if there is a single session.
- Added the ability to crop a selection from a partition into a new PsiStore.
(3) Improvements and updates in the data layer and visualization infrastructure:
- Refactored large parts of the class hierarchy for
VisualizationObject
andVisualizationObjectView
, to clarify distinctions across the dimensions of data access type (stream value or stream interval), data type, and view type. - Refactored the data access layer to better manage object lifetimes and handle disposable (and shared) objects without generating leaks.
- Refactored stream adapters to rely on virtual methods. Apart from a
GetAdaptedValue
virtual method that was introduced to handle the adaptation, aDispose(TDestination destination)
method has been introduced that allows the adapters to properly dispose of items they allocate in the process of adaptation. - Refactored
StreamTreeNode
for extensibility by creating a class hierarchy of different kinds of stream tree nodes, and moved functionality fromPluginMap
intoStreamTreeNode
and its related classes. - Updated
TimelineCanvasVisualizationObjectView
to be able to work with any stream visualization object, not only timeline visualization objects. This allows creation of an instant value visualization objects that use a timeline view to display information. - Removed Pool infrastructure from PsiStudio since it had become obsolete and was not being used.
(4) Other bug fixes and improvements:
- Fixed UI problem where it was necessary to click on the text of the stream tree node (rather than entire row) to activate context menu.
- Systematized displayed property names for datasets, sessions, partitions, streams.
- Fixed an issue where instant data was not immediately populated when opening a new layout.
- Fixed an issue that did not allow partition names to be changed via property editor.
- Fixed an issue whereby the partitions did not update correctly in the Datasets view after a batch processing task was run.
- Fixed an issue in
DisplayImage
whereby images were not immediately shown. - Fixed an issue whereby synchronization between a visualization object containing a derived stream member and the stream tree was failing.
- Eliminated this old InitNew() pattern we had in the visualization objects as it was no longer necessary.
- Updated infrastructure for computing the order in which visualizers are added to context menus.
- Fixed an issue in PsiStudio on refreshing the stores after running a batch processing task that overwrites and existing store.
- Fixed an issue in visualization which would lead to problems if the partition name was different from the underlying store name.
- Fixed an issue in
DisplayImage
where initialization would not happen on the UI thread. - Fixed an issue where users were unable to clear the selection markers if only one of the start or end markers had been set.
- Fixed an issue where PsiStudio could sometimes crash when loading a Dataset if one of the Partitions in the Dataset was invalid or missing.
- Fixed dark text rendering of expandable property names.
- Updated infrastructure for importers to allow better lifetime management of objects. The
Importer.OpenStream<>
,OpenStreamIndex<>
, andSessionImporter.OpenStream<>
methods now take allocator and deallocator optional parameters, which in turn enables using shared pools when reading streams of aShared<T>
type. The importer component now deallocates of shared objects after post via the provided deallocator (by default it disposes of Disposable objects). -
PsiStore.Open(...)
now accepts an optionalusePerStreamReaders
flag causing separate importers for each opened stream (default=true). This allows messages to be emitted at a pipeline time approximating the original creation time, regardless of physical interleaved ordering. Setting this parameter tofalse
reverts to the older behavior prior to this release. - Fixes for computations of pipeline progress estimates.
- Added an overload to UnmanagedBuffer.CopyFrom that takes an offset and length.
- Added
PsiStore.SummarizeDistinctKeysInSupplementalMetadata(...)
to add metadata enumerating known keys. - Forced PsiStore.EditStream operations to use Unlimited policy to make sure no messages are dropped.
- Changed Exporter.Write() and .WriteEnvelopes() method to take in IProducer, not just Emitter.
- Added Size and StreamCount members on IStreamReader interface, as well as Dataset/Session/Partition classes. (Breaking change).
- Added deliveryPolicy and enableDiagnostics optional parameters on the CreateDerivedPartition APIs.
- The stream metadata classes (starting with the
IStreamMetadata
interface) have been updated:-
MessageCount
changed fromint
tolong
. -
AverageLatencyMs
renamed toAverageMessageLatencyMs
and changed fromint
todouble
. -
AverageMessageSize
changed fromint
todouble
. -
PartitionName
andPartitionPath
renamed toStoreName
andStorePath
. -
FirstMessageTime
renamed toFirstMessageCreationTime
. -
LastMessageTime
renamed toLastMessageCreationTime
.
-
-
PsiStreamMetadata
changes (in addition toIStreamMetadata
):- Removed
AverageFrequency
. -
AverageLatencyMs
renamed toAverageMessageLatencyMs
and is now computed fromMessageCount
andLatencyCumulativeSum
, is now adouble
and is now in milliseconds rather than microseconds. -
AverageMessageSize
is now computed fromMessageCount
andMessageSizeCumulativeSum
and is now adouble
. -
MessageSizeCumulativeSum
added (long
). -
LatencyCumulativeSum
added (long
).
- Removed
-
Message.Time
renamedMessage.CreationTime
. - Added
IStreamMetadata.StreamTimeInterval
(opened to closed time). - Added
StreamTimeInterval
property toImporter
,SessionImporter
andMetadataCache
. - Renamed
IStreamMetadata.AverageLatencyMs
toAverageMessageLatencyMs
. - Fixed an issue in which streams read from a \psi stores could produce messages with a
CreationTime
after the current pipeline time. - Fixed a store seek bug on index boundaries.
- Expanded and systematized computation of diagnostic statistics to include counts of emitted messages, as well as emitting and receiving latency. Updated the computation of some diagnostic metrics. Statistics now include both averages and last values. Renamed some of the other members.
- Components can customize the node name in diagnostic visualization by overriding the ToString() method on the component class.
- Fixed a bug on displaying stream information in the pipeline graph when two components are connected by multiple streams.
- Changed the default for heatmap visualization to be the receiving latency.
- Added an overload to
ImageBase.CopyFrom
that takes an offset and length. - Added image operators for color averaging.
- Added support for specifying invalid values in
PseudoColorize
image operation and stream operator. The pseudocolorized images are now in BGRA_32bpp format. - Added new RGB_24bpp PixelFormat.
- Changed
SetPixel
onDepthImages
to use ushort rather than int. - Added
GetPixel
andToPixelFormat
methods. - Added
GetPixelRange
method forDepthImage
. - Fixed incorrect byte ordering in
SetPixel
for certain pixel formats. - Fixed "Buffer too small" error when creating certain images with non-aligned strides.
- Fixed bugs in the
Flip
,Invert
andThreshold
operators. - Fixed several bugs related to pixel format conversions.
- Small fix in image decoder and encoder components to allow pass-through of null messages.
- Batch processing task attributes now allow for specifying the output store name and path, the replay descriptor, the pipeline delivery policy, and whether to enable pipeline diagnostics.
- Dataset APIs extended to allow for specifying pipeline delivery policy and enabling diagnostics when computing derived partitions.
- Improved progress estimates for computing derived partitions over datasets by considering session durations.
- Expanded progress window when running batch processing tasks to include data size, time elapsed and estimated time remaining.
- BREAKING CHANGE: Calibration: Merged
DepthExtensions
andRotationExtensions
into a new static class,CalibrationExtensions
.
- Added
OutputPath
parameter toConcat
,Crop
,Encode
operations. These operations are now outputting stores in the specified paths. If a store already exists at the specified output path, the tool asks for confirmation to overwrite. - Systematized the options and their explanations for the various commands.
- Added
analyze
verb to PsiStoreTool to measure how well messages streams are interleaved within a store.
- Fixed a bug in the bounding box computation of the TinyYoloV2 ONNX model runner.
This release includes a new ONNX model runner for ImageNet models, new components for screen and window capture, updates to annotation editing in Psistudio, as well as a number of bug fixes and updates.
- Added VisualCapture component to capture a System.Windows.Media.Visual to a stream of Shared
.
- Added WindowCapture component to capture a window specified by a window handle (defaulting primary screen) to a stream of Shared
.
- Decode MJPEG frames in Linux
MediaCapture
component. - Bug fix to avoid serializing/persisting entire underlying
MemoryStream
buffer inEncodedImage
. - Fixed bug where
FromImage()
would crash since it doesn't support 8bpp.
- Multiple updates to annotation editing. Fixed bug where users could not drop selection markers on top of an existing annotation. If a selection marker is to be dropped "close" to an existing annotation edge, PsiStudio now drops the marker exactly on the edge so that a collection of contiguous annotations can be easily created. Users can now directly edit an annotation in place; the text of an unrestricted annotation can be edited directly in the annotation itself, and clicking on a finite annotation opens a context menu where the user can immediately select the required value from a list. The properties of a newly created annotation are now automatically displayed. PsiStudio now displays an informative error message when a user attempts but is unable to place an annotation, for example, if the new annotation would overlap an existing annotation.
- Cleanly handle mismatched types when deserializing streams of custom types. Previously, when the current structure of some type T was different to the structure of T in an existing stream, PsiStudio would crash when it attempted to deserialize that stream. PsiStudio now displays a message indicating how the structure of T has changed and will stop attempting to read from the stream.
- Fixed bug in context menu where users had the option to display stream latencies in instant visualization panels. Latency plots should only be displayed in timeline panels.
- When snap-to-stream is enabled, PsiStudio now no longer attempts to snap to messages that are outside the current viewport.
- Bug fix where adding, removing and re-adding an image visualizer would crash PsiStudio.
- Bug fix: The current value of an instant visualization object is now cleared when it becomes unbound. Previously the last value (e.g. image) would continue to be displayed after the stream became unbound.
- Added visualization adapters for 3D Points, Lines, and Rectangles.
- Bug fix where saving annotations in PsiStudio would occasionally crash.
- Fixed bug where clicking on a timeline panel occasionally caused the entire timeline to shift by a large amount.
- Replay speed can now be reduced to a slower speed than 1.0. Clicking the down button when the speed is 1.0 or less will halve the current playback speed.
- Workaround for cache misses when visualizing very large stores. If the required data has not yet been read from disk PsiStudio now explicitly requests it rather than waiting for it to be loaded naturally.
- The states of the timing buttons in the toolbar are now preserved when switching layouts.
- Fixed bug where loading layout at startup sometimes crashed.
- Bug fix for correctly handling Nullable in the Expand Members context menu item for streams.
- Fixed several layout bugs in the diagnostics visualizer.
- Added support for Wired Hardware Sync for Azure Kinect Sensors.
- Changed default
TemporalSmoothing
parameter inAzureKinectBodyTrackerConfiguration
to 0.
- Bug fix for possible truncation and corruption of values in
StreamEnumerable<T>
. - Bug fix for potential delay in
RelativeTimeWindow
stream operator. - Removed obsolete
Store
class. - The
StreamMetadataBase
class has been moved from theMicrosoft.Psi.Audio
namespace to theMicrosoft.Psi
namespace.
- Added ONNX model runner for a variety of ImageNet models.
- Added QUICKSTART help topic for LINUX.
- Removed sample projects - These have moved to a separate Psi Samples repository.
- Moved api docs to DocFX and added DocFX build project.
- Added
encode
verb to PsiStoreTool to support JPEG encoding ofShared<Image>
.
The release brings several important updates to Platform for Situated Intelligence Studio (PsiStudio), including support for temporal data annotations, additional visualizers for cartesian charts and histograms, general updates for various UI elements, as well as bug fixes and other improvements.
At the runtime and infrastructure level, this release adds support for writing custom store readers (via the IStreamReader
interface) that can provide streaming data to \psi pipelines as well as in PsiStudio from data sources stored in third-party formats. As an example, we have written a WaveFileStreamReader
that allows reading data and visualizing data from WAVE files.
Finally, the release also introduces new \psi components that enable running Onnx machine learning models.
The full, more detailed list of changes is described below:
- Added support for temporal annotations. Users can now create streams of multi-track annotations based on custom annotation schemas.
- Added support for visualizing a property or field of a stream. Right-clicking on a stream and selecting “Expand members” will display these sub-fields and they can be visualized as well.
- Added initial support for dynamically loading batch processing tasks from 3rd party assemblies. The tasks are identified by the
[BatchProcessingTask]
attribute, and surfaced in the context menus for Dataset and Sessions. Currently only tasks that have a signature taking a (Pipeline
,SessionImporter
,Exporter
) are surfaced. - Added support for opening custom, third-party stores based on the available stream readers (e.g. choose
.wav
in file dialog). - Temporal selection region is now highlighted.
- Added new visualizers for camera intrinsics (visualizes a camera frustum), depth image point cloud, depth image mesh, and image with intrinsics.
- Added new visualizers for cartesian charts and histograms using LiveCharts.
-
CoordinateSystem
visualizer now supports size and diameter as separate properties. -
KinectBody
andAzureKinectBody
visualizers now support individual bone thickness and joint radius. - Panels now have a
BackgroundColor
property. - 2D and 3D panels now have a
RelativeWidth
property that allows for variable width instant panels in a container. - The layout selection combobox and associated buttons have been moved from the main toolbar to the Visualizations tab. Multiple minor bug fixes around switching layouts.
- Several engineering, refactoring, and code clean-up updates, including refactoring visualization for depth images, restructuring context-menu generation, rationalizing names for the
VisualizationObject
attribute, increased readability of type names the property browser. - Several additional bug fixes in
DepthExtensions.ProjectTo3D
,IntersectLineWithDepthMesh
,Point3D
visualizer, message visualization object. - BREAKING CHANGE: The
Microsoft.Psi.Visualization.Common
namespace has been renamedMicrosoft.Psi.Visualization
and the project Microsoft.Psi.Visualization.Common.Windows has been renamed Microsoft.Psi.Visualization.Windows.
- Added a new set of packages, Microsoft.Psi.Onnx.Gpu and Microsoft.Psi.Onnx.Cpu containing a component called
OnnxModelRunner
that enables running a specified Onnx model. - Added a new set of packages, Microsoft.Psi.Onnx.ModelRunners.Gpu and Microsoft.Psi.Onnx.ModelRunners.Cpu, containing wrapper components that enable more easily running specific Onnx models. As a first example,
TinyYoloV2OnnxModelRunner
enables running object detection over images with a TinyYoloV2 model. We plan to extend the set of model runners in the future.
- Added
IStreamReader
andIStreamMetadata
interfaces to support reading streams from custom storage formats. -
PsiStoreReader
is an implementation ofIStreamReader
for \psi stores. -
WaveFileStreamReader
is an implementation ofIStreamReader
over WAVE files. - Added extension methods to add sessions and partitions specific to \psi stores and WAVE file stores.
BREAKING CHANGES:
-
Importer
,Store
,Dataset
,Session
andPartition
APIs have changed to accommodate arbitrary stream readers. - APIs returning
PsiStreamMetadata
now returnIStreamMetadata
. - Static APIs specific to \psi stores have moved from
Store
toPsiStore
(Store
APIs are obsolete). -
Importer
constructor no longer takes name and path; instead it requires anIStreamReader
. -
PsiStore.Create(...)
now returnsPsiExporter
(baseExporter
is now abstract).
- Fields with the
[NonSerialized]
attribute are no longer skipped when cloning. To skip cloning such fields, register the containing type with theSkipNonSerializedFields
cloning flag. - Improved serialization compatibility for some .NET types across frameworks.
- Consolidated overloads for Pipeline.Create() factory method into a single method.
BREAKING CHANGES:
- Renamed
PsiStreamMetadata
andImporter
properties:-
Lifetime
->StreamTimeInterval
-
ActiveLifetime
->MessageCreationTimeInterval
-
OriginatingLifetime
->MessageOriginatingTimeInterval
-
FirstMessageTime
->FirstMessageCreationTime
-
LastMessageTime
->LastMessageCreationTime
-
- Renamed
IndexEntry.Time
->IndexEntry.CreationTime
. - Renamed
Envelope.Time
->Envelope.CreationTime
. - Cloning objects containing unmanaged pointers and
IntPtr
fields (e.g. delegates) is no longer allowed by default. This behavior may be overridden by registering the containing type using theKnownSerializers.Register
method with the cloning flagsClonePointerFields
andCloneIntPtrFields
respectively.
- Added
Configuration
member toAzureKinectSensor
.
- Checks on Unmanagedbuffer sizes to only copy as much data as is in the actual image (height*stride).
- Fixes to resize/scale to preserve sizes.
- BREAKING CHANGE: Changed
Rotate
andDetermineRotateWidthHeight
to support loose/tight rotation.
- BREAKING CHANGE: The
OutputFormat
parameter inAudioCaptureConfiguration
has been renamedFormat
. - BREAKING CHANGE: The
InputFormat
parameter inAudioPlayerConfiguration
has been renamedFormat
.
- BREAKING CHANGE: Renamed
[Task]
attribute used to define tasks to[BatchProcessingTask]
. - BREAKING CHANGE: Removed
SimpleVoiceActivityDetector
component. - Dataset processing APIs now support progress reporting.
- Fixed bug in
ImageAnalyzer
component where region was being ignored. - Set auto CRLF handling for the whole repository.
- Updated symbol packages to the .snupkg format.
The major changes in this release include the addition of support for the Azure Kinect sensor (including body tracking), an overhaul of Microsoft.Psi.Imaging, some additional Kinect for Windows component cleanup, visualization updates, as well as an upgrade of projects to Visual Studio 2019 and .NET Core 3.1. Miscellaneous updates were also made to other areas (see more details below).
A Note About Coordinate Systems: This release begins to rationalize the handling of coordinate system bases throughout the code base. We adopt the MathNet.Spatial convention everywhere, in which the X-axis represents "forward", the Y-axis represents "left", and the Z-axis represents "up". Matrix representations assume column vectors and are stored in column-major order. These changes particularly affect code and data from Microsoft.Psi.Calibration as well as the Kinect and Azure Kinect sensor and body joint representations.
- Added
AzureKinectSensor
andAzureKinectBodyTracker
components that enable using the Azure Kinect sensor with \psi (see more: Overview, Sample, Azure Kinect Documentation). - The Kinect and AzureKinect body visualizers have moved into their own projects and need to be loaded as 3rd party visualizers for PsiStudio to become available.
A number of updates were made to the imaging infrastructure in Microsoft.Psi.Imaging, including the addition of support for depth images.
BREAKING CHANGES:
-
EncodedImage
now requireswidth
,height
andpixelFormat
to be specified on construction. -
EncodedImagePool
is now keyed, andGetOrCreate()
now requires the encoded imagewidth
,height
andpixelFormat
. - Renamed
Image.FromManagedImage()
toImage.FromBitmap()
. - Renamed
Image.ToManagedImage()
toImage.ToBitmap()
. - Renamed
ImagePool.GetOrCreate(Bitmap)
toImagePool.GetOrCreateFromBitmap(Bitmap)
. -
Image.EncodeFrom()
now takes in anIImageToStreamEncoder
instead of anAction<>
. - Moved
ImageEncoder
andImageDecoder
\psi components from the platform-specific .Windows and .Linux projects into the cross-platform Microsoft.Psi.Imaging project. These components are now provided anIImageToStreamEncoder
orIImageFromStreamDecoder
object to do the actual encoding and decoding. A variety of implementations (e.g. for JPEG, PNG) are available in the corresponding platform-specific .Windows and .Linux projects. - Removed the static method
ImageEncoder.EncodeFrom(encoder, ...)
and replaced withencoder.EncodeFrom(...)
. - Renamed
CompressionMethod.PNG
toCompressionMethod.Png
. - Renamed the
Convert()
stream operator for shared images toToPixelFormat()
, for clarity and to match the underlying component. - The 3D related operators from
DepthExtensions
in Calibration now operate only on depth images. -
KinectSensor.DepthImage
stream now returns a stream ofDepthImage
rather than a stream ofImage
. - The
Window()
operator now correctly handlesTimeInterval
inclusivity rather than excluding messages on the boundary. Additionally, the operator no longer accepts awaitForCompleteWindow
parameter and never delays the initial window.
Other changes:
- Added new
DepthImage
andEncodedDepthImage
types to represent depth images, and a set of associated operators. - Overhauled the various image processing functions and corresponding stream operators to rationalize the API.
- Added support for preserving the pixel format when decoding an
EncodedImage
. - Fixed a number of bugs, including issues related to pixel formats and decoding.
- Kinect body visualization removed from Visualization.Common and put in its own project so that Visualization.Common and PsiStudio no longer have a dependency on the Kinect SDK.
-
KinectDepth
visualizer is now a more generalDepthImage
visualizer supporting both mesh and point cloud views. - Updated handling of Shared objects in instant visualizers.
- PsiStudio now starts up with no default dataset and the main window titlebar now displays the name of the current Dataset and the name of the stream currently being snapped to (if any).
- Updated visualization object selection algorithm to select the visualizer the user most likely wants to use.
- A StreamVisualizationObject can now only be snapped to when it is bound to a source, and if it becomes unbound then we remove snap-to-stream.
- A TimelineVisualizationObject can now only be zoomed to when it is bound to a source.
- Multiple bug fixes to StreamBinding serialization, we now write out the assembly qualified name of StreamAdapterType, and we now correctly serialize child objects of 3D visualization object collections.
- Code cleanup of XyzVisualizationPanel and its view including adding bidirectional properties for camera position, direction and field-of-view.
- Added support for specifying which items in an UpdatableModelVisual3DVisualizationObjectDictionary are visible via a predicate.
- Added PsiStudio cursor nudging feature.
- The pipeline visualizer now shows a dotted edge connection between subpipelines when they contain a connection (perhaps even between deep descendants) spanning them.
- Fixed bug where PsiStudio settings were sometimes not being correctly written when shutting down PsiStudio.
- Fixed bug involving the index view in StreamReaders for InstantVisualizationObjects when the user changed the value of CursorEpsilon.
- Changes to improve backward compatibility with some older stores.
- Added
BridgeTo
operator toProducer<T>
creating a connector as needed to bridge pipelines. - Raise
PipelineRun
event for all subpipelines before starting. - Change pipeline state to
Running
as soon as subpipelines have started. - Fixed a pipeline shutdown race condition.
- Fixed
DeliveryPolicy<T>.WithGuarantees
to do a union between guaranteed message sets.
- BREAKING CHANGE: Removed
Exporter.WriteEmitters
. - Added
CropInPlace
store operation. - Added support for supplemental metadata on streams.
- Fixed a bug in
Store.Copy
andStore.Crop
which were using theActiveTimeInterval
instead of theOriginatingTimeInterval
.
- Added "Pose" fields for
DepthDeviceCalibrationInfo
objects, which are the inverse of Extrinsics. - Added ability to specify distort/undistort as either normal or reversed. Normal would be the closed form equation which returns distorted points and the reverse version returns undistorted points.
- Added support for showing stream size information in PsiStoreTool.
- Added verb for removing a stream from a store to PsiStoreTool.
- Updated
Sample
operator with a new default, that allows it to sample the nearest message to the clock when no tolerance or relative-time-interval is specified. - The
Zip
component and corresponding operators were changed to returnT[]
.
- BREAKING CHANGES to
KinectSensor
component:- Removed custom Matrix code in
KinectInternalCalibration
and re-implemented using MathNet. - Removed
IKinectSensor
,KinectExtensions
,DepthExtensions
and all GZip encoding functionality. - Removed
DepthToColorConverter
and re-implemented as a generic imaging operator. - Moved Orientation functions and LevenbergMarquardt optimization into
Microsoft.Psi.Calibration
. - Changed
KinectBody
definition to store Joint poses as MathNet CoordinateSystems, with immediate conversion into MathNet basis (X Forward, Y Left, Z Up).
- Removed custom Matrix code in
- Added a constructor for
AudioCapture
component allowing an output format to be specified. - Added ALSA error codes in exceptions thrown by the Linux
AudioCapture
andAudioSource
components. - Fixed Linux audio and video components to work on 32-bit OS.
- Converted
RealSenseSensor
component to use newDepthImage
. - Added a constructor for
MediaSource
component that accepts aSystem.IO.Stream
as input. - Fixed an issue with
Mpeg4Writer
where image data was being written in RGB instead of BGR.
- This release moves .NET Core projects to version 3.1. As a result, Visual Studio 2019 is now required to build the code. Building with Visual Studio 2017 will no longer be supported.
- Added code analysis to projects.
- The following NuGet packages were updated to the versions shown:
- Extended.Wpf.Toolkit (3.6.0)
- HelixToolkit.Wpf (2.11.0)
- MathNet.Numerics.Signed (4.9.1)
- MathNet.Spatial.Signed (0.6.0)
- MessagePack (2.1.90)
- Microsoft.NET.Test.Sdk (16.6.1)
- MSTest.TestAdapter (2.1.1)
- MSTest.TestFramework (2.1.1)
- System.ServiceModel.Primitives (4.7.0)
- Various bug fixes, typographical corrections, and code cleanup.
This update brings several new developments, and numerous improvements and bug fixes in the \psi Runtime, Tools and Components.
- In Runtime the major changes include the addition of new types of delivery policies, emitter validators, and streamlining of the APIs for running the pipelines (see more details below)
- In Tools a number of significant updates have been brought to PsiStudio, including performance improvements, support for 3rd party visualizers, and matrix layouting, and other extended functionality. Similarly, the performance of the Diagnostics subsystem has been improved and visualization functionality for diagnostics data has been extended. A number of updates have been made also to PsiStoreTool.
- In Stream Operators and Components a number of extensions and bug fixes have been added.
A number changes have been made to clean up and streamline the APIs for pipeline execution. The changes below are breaking changes:
- Removed the
Pipeline.Run(TimeSpan)
method as its meaning was inconsistent with running a pipeline with a finite originating time interval. To run a pipeline for a given wall-clock duration, callRunAsync()
, followed byWaitAll(TimeSpan)
, then dispose the pipeline to stop it. -
ReplayDescriptors
no longer define theUseOriginatingTime
property as it was redundant. We always assume originating times when reading messages within a given time interval from a store. Consequently,Pipeline.ProposeReplayTime()
now takes only a singleTimeInterval
parameter representing the originating time interval. -
Pipeline.WaitAny
has been removed. To be notified of source component completion, use thePipeline.ComponentCompleted
event instead. - Pipeline replay at speeds other than real-time or maximum speed are no longer supported. Consequently, the
replaySpeedFactor
parameter andReplaySpeedFactor
property have been removed thePipeline.Run
methods and theReplayDescriptor
class respectively. - The
deliveryPolicy
parameter for the Pipeline and Subpipeline constructors has been renamed todefaultDeliveryPolicy
.
The available set of delivery policies has been extended. See more information in Delivery Policies tutorial:
- Added support for throttling policies via DeliveryPolicy.Throttled.
- Added a new lossless policy that attempts synchronous delivery or throttles otherwise via DeliveryPolicy.SynchronousOrThrottle.
- Added support for data-driven delivery policies via typed delivery policies with guarantees.
A number of additional changes have been made to the Runtime:
-
Pipeline.RunAsync
now supports progress reporting by means of anIProgress<double>
object. - The pipeline start time is now captured in a new
Pipeline.StartTime
property. - Subpipelines provide access to the parent pipeline via the protected
ParentPipeline
property. - When persisting a stream to a store, the stream closing time is now saved in the store catalog.
- Emitters now support optional user-defined message validation.
- Fixed a bug where the presence of an infinite source component within a subpipeline was in some cases causing the pipeline to terminate.
- Added message validation for emitters, and support for pipeline level defaults.
A number of significant updates have been brought to PsiStudio, including performance improvements, support for 3rd party visualizers, and matrix layouting, and other extended functionality:
Performance updates:
- Improved performance when reading data from disk for instant visualization objects. Previously these reads were occurring on the UI thread whereas now they happen in a worker thread. Now they are optimized so that the same piece of data is not read multiple times for different visualization objects.
New functionality and bug fixes:
- In order to make better use of available screen real estate, instant visualization objects (2D and 3D visualization objects) can now be arranged in a horizontal matrix in the visualization container. These matrices can contain from one to five instant visualization objects.
- PsiStudio can now visualize the Messages and Latencies for all streams, regardless of type.
- The states of the timing info buttons are now persisted across PsiStudio sessions.
- When the use clicks in a visualization panel, PsiStudio displays the properties of the first visualization object in that panel, or if the panel has no visualization objects it displays the properties of the panel itself. Similarly, whenever the user adds a new visualization object to the visualization container its properties are immediately displayed.
- All DateTime displays in PsiStudio now show the value down to 1/10,000 of a second.
- Fixed bug where certain line segments were not drawn if the value of any of the data points was NaN or positive infinity or negative infinity.
- Users can now play multiple audio streams simultaneously.
- Added snap to stream context menu item to the timeline panel.
- 3D visualization panel now supports camera manipulation and the execution of camera animation storyboards.
Support for 3rd party visualizers:
- Added support for 3rd party visualizers. Users who have written their own visualizers can now use them in PsiStudio in a much more seamless manner. Add a new node called
<AdditionalAssemblies>
to the PisStudio settings file atMyDocuments/PsiStudio/PsiStudioSettings.xml
. Any visualization object marked with the[VisualizationObject]
attribute, and any stream adapter marked with the[StreamAdapter]
attribute will be automatically added to PsiStudio’s list of visualizers.
Other updates:
- For historical reasons related to PsiStudio previously supporting a COM model, Visualization Objects and Visualization Panels each had a separate
Configuration
class. This configuration has now been merged back into the visualization objects or visualization panels, simplifying the code. - Updated the 3D visualization objects to have a more hierarchial structure.
- Layout serialization now includes a version number to help prevent loading a layout that is not compatible with the current version of PsiStudio.
The diagnostics system enables collecting and visualizing application diagnostics in PsiStudio. This release includes many performance improvements to the diagnostics system. The overhead of collecting diagnostics information has been improved. Message sizes are half what they were and rendering in PsiStudio has been improved by orders of magnitude. Many bug fixes have been made as well, mosting having to do with visualization in PsiStudio. There have been breaking changes to the diagnostics message type affecting visualization.
Visualization:
- To simplify the graph display, we now allow checking a box Show Exporter Links to show/hide connectors and edges from them to Exporters.
- Connectors render with a little icon (☍) and
Joins
now render in their own color with a simple plus (+). - Type names have been simplified from their fully qualified form to a more natural “code form” such as
(List<string>, Dictionary<char, (int[], double)>)
. - Connector tooltip now shows target/source.
- Emitter names can now be optionally shown on edge labels.
- Delivery policies can now be optionally shown on edge labels and hilighting of edges by delivery policy is now supported.
- Property panel now allows toggling emitter/receiver names as well as delivery policies.
- A dotted line edge is now shown indicating a connector into a descendant pipeline. That is, an edge to the direct child containing the descendant.
Naming Improvements:
-
Subpipeline
andExporter
node label is now the pipeline name (with type name now as a tooltip) - Components can now provide names via
ToString()
used to render nodes (rather than the default type name)
Miscellaneous:
- Averaging is performed by unit time rather than past n-messages and added dropped & processed averages.
- Added
concat
verb to PsiStoreTool (as well asStore.Concat(…)
static method) allowing concatenation of stores. - Added
crop
verb to PsiStoreTool allowing cropping of stores (exposingStore.Crop(…)
functionality). - Added
Microsoft.Psi.TaskAttribute
to mark “task” methods as light-weight jobs that may be executed by PsiStoreTool (and future PsiStudio). - Added
tasks
andexec
verbs toPsiStoreTool
allowing listing and executing tasks.
A number of BREAKING CHANGES were made to stream operators:
-
Windowing operators now produce streams of
T[]
rather thanIEnumerable<T>
, because the latter are not serializable. - The mathematical and statistical operators no longer apply to streams of
IEnumerable<T>
, but to streams ofT[]
. - Removed the
Sequence<T>
operator overloads that take anIEnumerator
argument. Use the overloads that takes anIEnumerable
instead. - Removed the optional
interval
andalignmentDateTime
parameters from theOnce
andReturn
operators.
A number of other additional enhancements were made to other stream operators:
- Added support for default delivery policies and naming of the
Parallel
composite components and operators. - Generator changes:
- Streams produced by
Generators
are now aligned by default to the pipeline start time, unless explicit message times are supplied. - Fixed a bug in the time alignment computation which may in rare cases result in a missing first message when an
alignmentDateTime
is supplied. - Fixed a bug where some
Generators
were not obeying aReplayDescriptor
with a finite end time.
- Streams produced by
- Added a
DynamicWindow
component and corresponding (Window(…)
overload) stream operators that enable data-driven windowing. - Small extension (with optional parameter) to
RelativeTimeWindow
component and correspondingWindow()
stream operator, enabling them to post results immediately before seeing a full window. Default behavior is still to wait for the complete window. - Added
Merge
operator, which trivially combines one or more streams of typeT
into a single stream of typeMessage<T>
, each with generated at the pipeline time when the message arrives atMerge
, but containing the original envelope. - Added
Zip
operator, which operates likeMerge
but ensures delivery in originating time order (ordered within single tick by stream ID) - Added
First(n)
operator, which filters a stream to the first n messages. - Added a
WriteEnvelopes
method for streams that enables writing only the envelopes from a stream (without the payload) to the store. -
Importer.OpenStream<T>
now checks that the type of the underlying stream messages matches the supplied type argumentT
and throws an exception if they do not match. - Performance improvements to the
Std
operator. - Minor changes in the way
Generator
,Importer
andExporter
apply delivery policies.
A number of changes were made in a variety of components:
- AudioCapture component
- Component now uses WASAPI event-driven capture by default. The previous behavior of pulling data from the engine at the target latency is still available by setting the UseEventDrivenCapture configuration parameter to false.
- The default audio capture buffer size may now be overridden by setting the
AudioEngineBuffer
configuration parameter. Larger values may help reduce the likelihood of audio glitches. - Component now throws an
IOException
if it is unable to create the audio capture device.
- AzureSpeechRecognizer component:
- Final recognition results are now aligned in originating time to the input voice activity detection signal.
- SystemSpeechSynthesizer component:
- Added
ReceiveSsml
andCancelAll
receivers to support issuing SSML-based speech synthesis requests and stopping of speech synthesis.
- Added
- SystemVoiceActivityDetection component:
- Added
InitialSilenceTimeoutMs
,BabbleTimeoutMs
,EndSilenceTimeoutMs
, andEndSilenceTimeoutAmbiguousMs
parameters in the component configuration. More details about these parameters may be found in the SpeechRecognitionEngine documentation.
- Added
- Face Identification:
- The
FaceRecognizer
component, which identifies faces in a stream ofShared<Image>
, has been updated to the latest Azure Face APIs. - A
PersonGroup
class supports managing person groups in Azure (creating, adding faces, training, deletion). - Also included are
PsiStoreTool
tasks supporting creating, training and testing from a directory of labeled static image files.
- The
-
Microsoft.Psi.Calibration
changes (these are Breaking Changes):- Added support for full rational distortion model (with k1-k6 parameters)
- Calibration representations now uses a canonical coordinate system throughout, in alignment with MathNet (right-handed, X forward, Y Left, Z up, column-vectors).
- Moved
KinectCameraCalibration
class and corresponding interface intoMicrosoft.Psi.Calibration
and renamed toDepthDeviceCalibrationInfo
-
SystemCalibration
was renamed toMultiCameraCalibration
- Moved
ProjectTo3D
class fromMicrosoft.Psi.Kinect
toMicrosoft.Psi.Calibration
.
- Added
Microsoft.Psi.DeviceManagement
project which enables device enumeration. UpdatedMicrosoft.Psi.Kinect
project to allow for enumerating all Kinect devices. - The
Microsoft.Psi.CognitiveServices.Vision.Windows
project is now .NET Standard and has been renamedMicrosoft.Psi.CognitiveServices.Vision
.
- Updated projects in solution to new .csproj style.
- Updated and consolidated NuGet references to the latest versions.
- Fixed a bug that was causing two unit tests to fail occasionally.
The major changes in this release are updates and added functionality in the stream operators for fusion and synchronization, operators for interpolation and sampling, and operators for generating streams. The release also includes a number of other updates and bug fixes - see more details below.
A number of changes have been made to the set of interpolators, and the synchronization operators (for the latest documentation on synchronization and stream fusion and merging see here):
- The set of interpolators has been restructured, generalized and expanded. Interpolators can now produce a result that has a different type than the input message type, and allow for matching with the
Nearest
,First
orLast
message. A taxonomy of greedy versus reproducible interpolators has been defined. - Added support for an adjacent-values-interpolator, and constructed a linear interpolator based on it.
- The existing
Join
operator that allows for reproducible stream fusion and correct synchronization now operates with reproducible interpolators. - A
Fuse
operator has been added to enable stream fusion with any generic interpolator (reproducible or greedy). - Interpolators now accept a user-specified default value to use.
A number of other operators for sampling, generating streams, and statistical computations have also been updated:
- Redesign of
Sample
andInterpolate
stream operators. Both operators are driven by a clock stream. TheSample
stream operator selects the nearest message to the sampling point (within a tolerance or relative time window). TheInterpolate
operator uses a specified interpolator to generate the result. The documentation for sampling and interpolation is available here. - The
Generators.*
methods have been updated to allow the developer to specify whether the generated streams should stay open or close after the enumeration terminates. The current documentation on stream generators is available here. - Refactoring and clean-up of mathematical and statistical operators.
- Added extension methods to provide access to graph-wide statistics such as total queued messages, total dropped, etc. across the whole graph.
- Enabing diagnostics at pipeline creation time (e.g.
Pipeline.Create(true, ...)
) used to take an optionalTimeSpan
interval. Now it takes an optionalDiagnosticsConfiguration
; containing this and other settings. - New
DiagnosticsConfiguration.TrackMessageSize
flag determines whether message byte counts are tracked (false
by default, as it is an expensive operation).
- Minor icon updates.
- Ensure that dialog boxes appear above the PsiStudio main window.
- Moved
EncodedImage
andEncodedImagePool
out ofMicrosoft.Psi.Imaging.Windows
and.Linux
into the central .NET Standard sharedMicrosoft.Psi.Imaging
. This is to enable referencing a singleEncodedImage
type across platforms. - Interop: Updated CSV format docs for new behavior.
- Added sensor_msgs/Image message type to the ROS bridge.
- Added more informative exception message when posting messages on a stream with out-of-order sequence IDs.
- Fixed bug in pipeline diagnostics capture regarding counts of messages dropped.
- Fixed bug in queue-size-constrained delivery policies (including
DeliveryPolicy.Latest
) which sometimes led to a queue size larger by one than the max, and allowed additional messages to be processed.
- As a result of the re-organization of interpolators, the
Match
static class and the interpolators it contained have been renamed. Specifically:Match.Best
->Reproducible.Nearest
,Match.Exact
->Reproducible.Exact
; similar for the OrDefault versions. -
Interpolator<T>
was replaced withInterpolator<TIn, TOut>
to support interpolation to a different output type. -
Generators.Sequence
,Generators.Repeat
andGenerators.Range
require specifying an explicit time interval (previously the time interval parameter was optional and defaulted to 1 tick). - The existing overloads of the
Sample
stream operator that used an interpolator were renamed toInterpolate
. - The OpenCV Sample now requires OpenCV 4.1.1 to compile. You must also set the environment variable OpenCVDir_V4 to point to your local installation of OpenCV 4.1.1.
- Enabing diagnostics (e.g.
Pipeline.Create(true, ...)
) now takes an optionalDiagnosticsConfiguration
rather than aTimeSpan
interval.
The main highlights in this release are:
- A major refresh of the
Platform for Situated Intelligence Studio
user interface, including a new color scheme, updated icons, and new functionality.
- The installation documentation has been updated to detail how to build the \psi codebase on Windows (using either Visual Studio 2017 or 2019), or on Linux.
- Updated the look and feel of the user interface, including new colors and icons.
- The
Datasets
treeview and theVisualizations
treeview now featureExpand All
andCollapse All
buttons to fully expand and collapse the tree. - The
Visualizations
treeview now contains a new button to synchronize between it and theDatasets
treeview. When a stream in theVisualizations
treeview is selected, clicking this button will select the source stream in theDatasets
treeveiw and bring it into view. This feature is useful for those times when you forget exactly which source stream a visualization is depicting. - When in
Live
cursor mode, the timing information for the selection start and end markers is now not displayed as this information is irrelevent when in live mode. - Added new context menu item
Zoom to Stream Extents
to streams in theVisualizations
tree view. - Added new context menu item
Zoom to Panel Extents
to panels in theVisualizations
tree view.
- Fixed a bug in the
Join
operator that made it not correctly take into account open relative time intervals. See this issue. - Fixed bug in the
Diagnostics
visualizer where connectors between subpipelines were being incorrectly rendered. - Fixed bug in the
Diagnostics
visualizer that would cause it to crash if it encountered messages with unserializable fields. We now return 0 for the size of such messages. - Fixed bug in
Rodrigues
method in theOrientation
class ofMicrosoft.Psi.Kinect.Windows
. This method was returning an NaN rotation matrix when passed a zero rotation vector, it now returns an identity matrix instead.
IMPORTANT NOTE:
In this release we have retargeted all .NET Framework projects to .NET Framework 4.7.2, and all native projects to Windows 10 SDK version 10.0.18362.0. In order to build the source code, please ensure that the following additional components are added to your Visual Studio installation:
If you are using Visual Studio 2019, you may install these components using the Visual Studio Installer, or via the download links provided above. For Visual Studio 2017, you may use the Visual Studio Installer to install the .NET Framework 4.7.2 targeting pack, but you will need to download and install the Windows 10 SDK (10.0.18362.0) using the provided link.
Some of the main highlights in this release include:
- Support for pipeline structure visualization, including message flow and diagnostics information (e.g. latencies, throughputs, stream statistics, etc.).
- Retargeted projects to .NET Framework 4.7.2 and Windows 10 SDK 10.0.18362.0.
- Updated NuGet package references in many projects to more recent versions.
- Runtime API changes and bug fixes.
- Moved documentation to wiki pages.
- Created gitter channel.
- The
ISourceComponent.Stop
method now takes afinalOriginatingTime
argument representing the pipeline shutdown time and anotifyCompleted
delegate which the component must now call to notify the pipeline that it has completed generating source messages up to thefinalOriginatingTime
. For more details, see here. - Pipeline exception handling is now exposed through new
PipelineException
event rather than overloading thePipelineCompleted
event to also serve as an exception handler. - Input and output connectors for subpipelines should now be constructed using the new
CreateInputConnectorFrom
andCreateOutputConnectorTo
static methods of theSubpipeline
class. - The
Match.Any
interpolator has been removed. - Enabing diagnostics at pipeline creation time (e.g.
Pipeline.Create(true, ...)
) used to take an optionalTimeSpan
interval. Now it takes an optionalDiagnosticsConfiguration
; containing this and other settings.
Several significant changes were made to the pipeline startup and shutdown logic:
- There is now a single scheduler for the main pipeline and all subpipelines. To facilitate management of work items in subpipelines, a new
SchedulerContext
has been introduced to which work items may be assigned. - Message delivery in a pipeline is now delayed at startup until all source components in the pipeline have been activated.
- To enable correct and reproducibile pipeline shutdown, changes were made to the
ISourceComponent
that will need to be implemented by stream sources. For more details, see here. - Renamed the
PipelineElement
Start
andStop
methods toActivate
andDeactivate
respectively. - Added a check in
Receiver.OnSubscribe
to ensure that emitters and receivers are not connected across pipelines (connectors should be used instead). - Changes to the pipeline exception handling mechanism via a new
PipelineException
event. - New
DiagnosticsConfiguration.TrackMessageSize
flag determines whether message byte counts are tracked (false
by default, as it is an expensive operation).
- Updated
ParallelSparse
andExporter
to use the composite component pattern, i.e. they are now implemented as a single subpipeline component. - Added
MessageConnector
andIConnector
interface to support connectors that wrap the entire message as they pass it along (used byExporter
). - Added new static methods
CreateInputConnectorFrom
andCreateOutputConnectorTo
onSubpipeline
for creating connectors between pipelines. - Changed the
StreamEnumerable
component (used by theToObservable
operator) to listen to theUnsubscribed
event of its source rather than thePipelineCompleted
event.
- New pipeline diagnostics visualization feature enables visualization of the pipeline structure and message flow statistics in PsiStudio.
- Added a new toolbar button that lets the user toggle whether the navigator's cursor follows the mouse cursor when in manual cursor mode. This setting can be toggled via the button and also via the shortcut "Alt+F" (enabled by default).
- When playing back in repeat/loop mode, we now ensure the cursor remains visible after looping from the end of the stream back to the beginning.
- Added support for visualizing streams of strings.
- Generic message visualizer now shows string representation of message in live legend.
- Added support for formatting in legends.
- Moved rectangle and coordinate system visualizers out of PsiStudio and into Visualization.Common.
- Created interface IView3D as well as a base version for handling collections of 3D visuals.
- Created a single base 3D visualization object, removing a lot of duplicate code.
- Fixed an issue which caused the pipeline to appear to stop responding if an exception is thrown in a subpipeline's attached
PipelineRun
event handler. - Fixed a bug in PsiStudio where the cursor remains a hand after dragging within a timeline visualization panel.
- Fixes to various visualizers (cleaned-up pattern for InitNew and handling config and property changes).
- Fixed an issue where live stores sometimes had no last message time.
- Fixed an AccessViolation related to
IAudioEndpointVolumeCallback
in the Windows audio components. - Fixed a bug in one of the tuple-flattening
Join
operator overloads which was causing it to use the incorrect interpolator. - Fixed a thread safety issue in the
Join
operator. - Fixed a thread cleanup issue on disposal in
InfiniteFileWriter
.
There are many additions and updates in this release, but the major changes can be summarized as:
- Improvements and fixes to the pipeline shutdown procedure, as well as fixes to the
Parallel
andJoin
operators to support reproducible dynamic sub-pipeline construction and teardown via theParallel
operator. - Streamlining the use of
Shared<T>
for more efficient messaging of large objects such as images, and added in-depth documentation on this topic. - PsiStudio now supports connecting to live stores and fast layout switching.
- The pipeline startup and shutdown procedure has been updated to support correct and reproducible shutdown where possible in dynamic pipelines. Specifically:
- Removed the
RegisterPipelineStart/Stop/Final
handlers. If the component implementsISourceComponent
, then use theStart
orStop
method, or use one of the following events instead:PipelineRun
,PipelineCompleted
,Receiver.Unsubscribed
. - The authoring of source components has been streamlined, and in the process the
IFiniteSourceComponent
andISourceComponent
interfaces have been unified into a single, newISourceComponent
interface. Implementers should call thenotifyCompletionTime
delegate supplied in the interface'sStart
method to indicate its completion time (orDateTime.MaxValue
if it is infinite).
- Removed the
- Removed
GroupBy
operator. -
MessagePack
no longer uses LZ4 compression during serialization, and now emitsDateTime
objects as Ticks rather than string representations.
There have been multiple changes made to the pipeline finalization code to support an orderly shutdown process.
- Subpipeline now has independent scheduler so that it may fully shut down independently of the main pipeline.
- Renamed several
Pipeline
events:-
PipelineCompletedEvent
event has been renamedPipelineCompleted
. -
ComponentCompletedEvent
event has been renamedComponentCompleted
. -
PipelineCompletionEventArgs
class has been renamedPipelineCompletedEventArgs
. - A new event
PipelineRun
has been added, it is raised whenPipeline.Run
(orPipeline.RunAsync
) has been called but before the components have started work.
-
-
Receiver
component now provides anUnsubscribed
event. -
Scheduler
now drops messages that were posted after the pipeline has shut down. -
Scheduler.Schedule
method now returns an indication of whether the call was ignored due to the pipeline having already shut down. - Bug fix in
ParallelSparse
so it works correctly with various branch termination policies. - Pipeline pause for quiescence now includes all child subpipeline schedulers.
- Fixed bug in the
ToEnumerable
operator where empty streams were not being handled correctly.
- Default branch termination policy now returns the OriginatingTime of the last message from that branch.
-
Parallel
operators now use aConnector
to bridge to Subpipelines. - Added a fix for Parallel with orDefault, which was not working properly in some cases due to errors in the matching function.
- Fixed bug in
Join
operator relating to the dynamic closing of secondary streams. - Fixed bug in
Match
component'sNearestMatch
function where if there were no available messages we would always returnInsufficientData
. We now first check if the stream is closed, and in this case we returnDoesNotExist
instead.
- The
Shared<T>.Create
method no longer supports specifying a SharedPool to which shared objects may be recycled. Recyclable shared objects must now be created from a shared pool. - The
SharedPool<T>.GetOrCreate
method no longer supports specifying a construction delegate, it must now be provided when theSharedPool<T>
is created.
- Added grammar-based intent detectors using
Microsoft.Speech
andSystem.Speech
- We now ensure that messages have monotonically increasing OriginatingTimes in the
MicrosoftSpeechRecognizer
component to comply with the new rule on theEmitter
component introduced in the last release. - Added support for
AzureSpeechServices
and marked as deprecated the obsolete componentBingSpeechRecognizer
. - The Speech event duration is now defined as the OriginatingTime of the last message where the VAD was still false (before it switched to true) to the last message where the VAD output was still true. Previously we were defining the event duration as the OriginatingTime of the first message where the VAD switched to true to the OriginatingTime of the first message where it went false again.
- Added
Reframe
operator toAudio
.
- Added new Psi component
PersonalityChat
to the Microsoft.Psi.CognitiveServices.Languages.Windows project that wraps Project Personality Chat.
- Removed all of the COM features that allowed a Psi application to launch and control PsiStudio. See next point.
- PsiStudio can now connect directly to the store of a running Psi application in the same way it connects to a previously created store. Users can seamlessly switch between Live mode and Playback mode when visualizing a live store.
- Simplified the hierarchy of
VisualizationObject
classes. - Added option to show tracking id in the
kinect body
visualizer. - When multiple audio streams are being visualized, users can now select which audio source should be played through the user's PC speakers by right-clicking the relevent stream in the
Visualizations
view. - Added new
Repeat
button to the toolbar so that playback will loop infinitely between theSelection Start
andSelection End
markers. - The position and layout of PsiStudio is now preserved when the application is shut down and is restored the next time PsiStudio is launched.
- Added
Layout Chooser
to the toolbar. This allows the user to quickly switch between different Visualization layouts. Users can create new layouts either from scratch or based on an existing layout and then save this new layout. Layouts are stored in/MyDocuments/PsiStudio/Layouts/
. - Loading and then switching between multiple sessions is now correctly supported. When a Visualization layout is selected, switching to a different session will automatically cause all Visualizers to rebind to the stores associated with this newly selected session.
- Added new icons to indicate when a
Visualization
has no stream to bind to in the current session.
- Fixed bug where the final extent of the store on disk was not being correctly truncated to the actual stream size after pipeline shutdown.
- Fixed bug in PsiStudio where zooming out a long way would cause the app to apparently freeze.
- Fixed bug where
StoreWriters
andStoreReaders
were generating different names for the same global mutex that indicated a store was currently live and being written to. - Fixed bug where replay time intervals in parent pipelines did not take into account the replay time intervals in subpipelines.
- Fixed memory leaks in
Microsoft.Psi.Imaging.Operators
class. - Improved performance of the serialization classes.
- Reduced startup delays in subpipelines.
- It is now a requirement that messages posted on an
Emitter
have strictly increasing originating times. Attempting to post multiple messages with the same originating time on the same stream will cause an exception to be thrown. - The
Buffer
,History
andWindow
operators have been unified as a single set ofWindow
operators which take either an index-based or a relative time-based interval. The index-based variants emit the initial buffer only after the total count of messages within the specified index interval have been accumulated, whereas the time-based variants emit the initial buffer as soon as messages within the specified relative time interval are available. - The following operators have been removed:
-
SelectMany
. -
Mirror
. -
Repeat
- usePair
instead. -
Buffer
- useWindow
instead. -
History
- useWindow
instead. Previous
-
-
Delivery Policies have been simplified and renamed:
- The
Throttled
policy has been removed. It will be re-introduced in a later release once issues around throttling have been resolved. - The
Default
,Immediate
andImmediateOrThrottle16
policies have been removed. - The
QueueSize
property has been renamedInitialQueueSize
. - The
MaximumLag
property has been renamedMaximumLatency
. - The
IsSynchronous
property has been renamedAttemptSynchronous
.
- The
- As a result of the above changes, many stream operators have been amended to take an optional
DeliveryPolicy
parameter. - The constructors for the
Connector
component and associatedCreateConnector
extension methods no longer take anowner
parameter. - The serialization format for
SystemCalibration
has changed:- The
ImageWidth
andImageHeight
properties have been removed. - The
NumberOfFrames
property has been moved to theCameraCalibration
class.
- The
- The
ICameraIntrinsics
interface has changed:- The signature of the
DistortPoint
method has changed.
- The signature of the
New Features in Platform for Situated Intelligence Studio:
- Updated the layout of the main Psi Studio screen. The Datasets tree view and the Visualizations tree view have been moved to the left hand side of the application. Furthermore they are no longer on separate tabs, they appear one above the other so that users no longer have to switch back and forth between tabs while laying out the visualizations. Datasets and Visualizations previously each had their own Properties pages, but now there is a single Properties page on the right hand side of the application that is able to display the properties of either type of object. Both the Datasets and Visualizations tree views and the Properties page can be resized or hidden completely to give more screen real estate to the main Visualization view.
- Added multi track event visualizer
TimeIntervalHistoryVisualizationObject
which is useful for visualizing multiple tracks of events having some finite timespan such as multiple speech-to-text streams. This visualizer will be loaded when visualizing streams containing messages of typeDictionary<string, List<(TimeInterval, string)>>
. TheDictionary
keys represent unique track IDs, each element in theList
represents an event to display in the track and contains a tuple of theTimeInterval
representing the start and end times of the event and a string representing the text that will be displayed inside the time interval. Note that since this is a history visualizer, each message should contain ALL events that have occurred up until the time of the message. This implies that the last message in the stream contains all of the data required to display the visualization. - Users can now visualize streams by dragging them from the Datasets tree view directly into the main Visualization panel.
- Added 'Snap to Stream' functionality on certain visualizers to snap the cursor to the messages of the snapped stream.
- Added 'Visualize Messages in New Panel' and 'Visualize Latency in New Panel' commands to the stream context menu.
- Psi Studio now automatically attempts to repair corrupted stores when opening them.
New Features in Platform for Situated Intelligence Runtime/Core:
- Initial version of data interop with the introduction of
Microsoft.Psi.Interop
, with support for MessagePack, JSON and CSV data formats and ZeroMQ transport. See the Interop topic for more details. - New
dynamic
store reader allows reading any stream from any store todynamic
primitives or toExpandoObject
ofdynamic
without requiring a reference to the .NET type of the stream messages. - New
PsiStoreTool
command-line tool which allows exploration of available streams in a store, conversion to other formats using interop, and saving to disk or sending over a message queue for consumption by other platforms and/or languages. - Exposed
Scheduler
as a parameter toPipeline
andClock
as a parameter toScheduler
. - Multiple handlers may now be registered on start, stop and final pipeline events.
- Improved
#TRACKLEAKS
debug information inRecyclingPool
.
New Features in Imaging:
- Added
SetPixel
method toImage
. - Added
DrawText
extension method forImage
. - Added support for
CameraIntrinsics
andCoordinateSystem
toSystemCalibration
. - The
IKinectCalibration
interface andKinectCalibration
class have been extended to support conversion from depth coordinates to color space coordinates using the newToColorSpace
operator.
Bug Fixes:
- Fixed several issues where visualization objects were not being displayed in the correct color in Psi Studio.
- Fixed a bug which would sometimes cause Psi Studio to crash when visualizing image streams.
- Fixed a bug which caused the mouse to move the cursor position during playback in Psi Studio.
- Fixed an issue causing streams to disappear at the end of playback in Psi Studio.
- Fixed a crash in Psi Studio when opening a layout created from a store which has since been deleted.
- Fixed a bug which sometimes caused timeline plots to be truncated when loading a layout in Psi Studio.
- Fixed an exception when closing Psi Studio after a live visualization session.
- Fixed a performance issue reading from Psi stores which occurred at the transition between consecutive data files.
- Fixed a bug where the
ImageCompressor
was not properly disposing of an encoded image after decoding it. - Fixed a bug where
ImagePool
would sometimes return a recycled image with incorrect dimensions. - Fixed a bug where the pipeline replay interval would sometimes extend beyond the lifetime of a stream being read from a store.
- Fixed a bug which caused
KinectSample
to crash when it detected no faces. - Fixed the
Scale
image extension method to throw an exception when attempting to call it on anImage
with an unsupported format. - Fixed a few intermittently failing unit tests.
- Fixed a bug which sometimes caused a loss of precision when computing the current pipeline time.
For this release we have added the compiler switch /Qspectre to our C++ projects which helps mitigate against Spectre security vulnerabilities. In order to successfully compile the Platform for Situated Intelligence solution you must upgrade your Visual Studio instance to version 15.7 or later and add the two following components to your Visual Studio installation:
- VC++ 2017 version version_number Libs for Spectre (x86 and x64)
- Visual C++ ATL (x86/x64) with Spectre Mitigations
For more information on the Spectre vulnerabilities and information about how to add the above components to Visual Studio, please see the following pages:
New Features in Platform for Situated Intelligence Studio:
- Added new PlotStyles to Timeline Visualizer. Plots can be rendered with the following styles:
- Direct (default): A straight line is drawn from each message to the next message to create a standard line plot.
- Step: Messages are joined by a horizontal line followed by a vertical line for visualizing quantized data.
- None: No lines are drawn between messages. If you select this plot style then you should also update the MarkerStyle from its default value of None or nothing will be drawn in the plot.
- Visualization Panels in Psi Studio can now be resized by dragging their bottom edge vertically.
- Visualization Panels can now also be re-ordered with the mouse via drag & drop.
- Users can now drag the visible portion of a Timeline Plot to the left or right using the mouse.
- Added modal window while loading a dataset to inform the user of the progress of the data load operation.
- Added new timing information toolbar buttons. These buttons can be used to display absolute message times, message times relative to the start of the session, and message times relative to the start of the current selection.
- Performance improvements when plotting messages.
New Features in Runtime:
- Join operator now matches against a final secondary message upon stream closing once it can be proven that no better match will exist.
- Added support for Parallel operators that take an Action rather than a Func.
- Adding components once a pipeline is running is no longer supported and now throws an exception. The recommended approach is to add a Subpipeline.
- Consolidated Windows SDK versions. Previously different parts of the toolset required different WinSDK versions, consolidated so that the only Windows SDK version that Psi now requires is 10.0.17134.0
- Improved how the pipeline shuts down to ensure that all existing messages are drained from it before stopping.
Bug Fixes:
- Fixed bug where switching Psi Studio from "Realtime" mode to "Playback" mode would result in the user being unable to move the timeline cursor or reposition the timeline.
- Fixed bug where we were trying to calculate the relative path of a partition to a dataset when the dataset was stored on the local disk but the partitions within the dataset were stored on a network share.
- Fixed bug where loading very large datasets would sometimes crash PsiStudio.
BREAKING CHANGES in this release:
- Removed several versions of the Parallel operator
- Renamed several components for consistency:
- AudioSource -> AudioCapture
- AudioSourceConfiguration -> AudioCaptureConfiguration
- AudioConfiguration (on linux) was eliminated and replaced by two corresponding classes AudioCaptureConfiguration and AudioPlayerConfiguration
- TransformImageComponent -> ImageTransformer
- AcousticFeatures -> AcousticFeaturesExtractor
- AcousticFeaturesConfiguration -> AcousticFeaturesExtractorConfiguration
- Renamed Parallel component to ParallelFixedLength to match naming convention with the other components.
- Partial speech recognition results for the SystemSpeechRecognizer, MicrosoftSpeechRecognizer and BingSpeechRecognizer components are now posted on a new stream named PartialRecognitionResults. The default Out stream now contains only final recognition results.
- Psi Studio will no longer load third party visualizers, for the time being it will only display its built-in visualizers.
Interim release with support for new devices, runtime enhancements and several API changes, as well as minor bug fixes:
- Added support for RealSense depth camera
- Added the
FFMPEGMediaSource
component for Linux - Added a [
Subpipeline
] class, enabling nested pipelines -
Parallel
now uses subpipelines -
Sequence
,Repeat
andRange
generators now allow time-aligned messages - Additional minor bug fixes.
Several API changes have been made:
-
Generators.Timer(...)
is nowTimers.Timer(...)
-
IStartable
has been replaced byISourceControl
/IFiniteSourceControl
and the way that components get notified about the pipeline starting and stopping has changed
Interim release with a few changes to the samples and some minor bug fixes:
- ArmControlROSSample is now RosArmControlSample.
- PsiRosTurtleSample is now RosTurtleSample.
- Added LinuxSpeechSample.
- KinectFaceDetector component now outputs an empty list if no face is detected.
- NuGet packages are now marked beta.
- Additional minor bug fixes.
Initial, beta version of the Platform for Situated Intelligence. Includes the Platform for Situated Intelligence runtime, visualization tools, and an initial set of components (mostly geared towards audio and visual processing). Relevant documents:
- Documentation - top-level documentation page for Platform for Situated Intelligence.
- Platform for Situated Intelligence Overview - high-level overview of the platform.
- NuGet Packages List - list of NuGet packages available in this release.
- Building the Code - information about how to build the code.
- Brief Introduction - brief introduction on how to write a simple application.
- Samples - list of samples available.
The Roadmap document provides insights about future planned developments.
- Basic Stream Operators
- Writing Components
- Pipeline Execution
- Delivery Policies
- Stream Fusion and Merging
- Interpolation and Sampling
- Windowing Operators
- Stream Generators
- Parallel Operator
- Intervals
- Data Visualization (PsiStudio)
- Data Annotation (PsiStudio)
- Distributed Systems
- Bridging to Other Ecosystems
- Debugging and Diagnostics
- Shared Objects
- Datasets
- Event Sources
- 3rd Party Visualizers
- 3rd Party Stream Readers
Components and Toolkits
- List of NuGet Packages
- List of Components
- Audio Overview
- Azure Kinect Overview
- Kinect Overview
- Speech and Language Overview
- Imaging Overview
- Media Overview
- ONNX Overview
- Finite State Machine Toolkit
- Mixed Reality Overview
- How to Build/Configure
- How to Define Tasks
- How to Place Holograms
- Data Types Collected
- System Transparency Note
Community
Project Management