Deep Object Pose Estimation (DOPE)

[lambdaprime]

DOPE (Deep Object Pose Estimation) is a system for Semantic Robotic Grasping of Household Objects described in the paper. Most of its resources (like model weights, datasets, training scripts etc) are available online.

dope4j is a Java interface to DOPE which is based on DJL.

• it exposes an API to DOPE network as well as provides Java based implementation of the DOPE original decoder. This allows users directly to use DOPE models in their Java applications without depending on ROS.
• based on dope4j API there is separately implemented standalone application dope4j-app which allows to experiment with DOPE. It allows to visualize data like belief maps, affinity fields with OpenCV. This is useful since allows to experiment with DOPE without need to publish images from camera or play rosbag. Instead users can provide path to local image files (with camera info) and see the results.
• to analyze the performance dope4j emits metrics. For that it relies on OpenTelemetry which allows users to chose between many observability backends. Metrics are disabled by default.

There are certain differences between Java and original Python decoder implementations:

• Decoding DOPE output is multistep process. Java decoder implementation supports Inspector which allows users to introspect every step of decoder and get access to belief maps, affinity fields, keypoints (separately vertices and center points) and finally bounding boxes.

• Based on Conformance to DOPE original results the deviation between results of two decoders is around 0.0999.