Datasets and models

[saed13]

ideas and analysis of the potential datasets and models

[Micropot]

• a2d2 datatset has 41k images already segmented with 38 different categories. We can use it for the recognition of the different elements of the street such as trafic lights or signs. Since this dataset as been created with real images, it's closer to the reality than a simulator, but it might not have any weird situations.
• KITTI Dataset is a dataset similar to the a2d2 dataset but well known for the training of AI on the autonomous driving subject. On their website, they have lost of benchmark on lot of different type of data collected my a car with cameras on it.

[jzzzza]

a2d2 (Audi Autonomous Driving Dataset)
    Task: Autonomous driving
    Pros:
        High-resolution camera images and point clouds
        Accurate sensor data and ground truth annotations
        Data captured in diverse weather conditions and scenarios
    Cons:
        Limited to Audi vehicles and their specific sensors

nuScenes
    Task: Autonomous driving
    Pros:
        Multi-modal dataset with camera, LiDAR, and RADAR data
        Large dataset with various urban driving scenarios
        Detailed annotations and object tracking
    Cons:
        Complex dataset structure
        Requires a significant amount of storage space and computational resources

BDD100k (Berkeley DeepDrive)
    Task: Autonomous driving, scene understanding, and object detection
    Pros:
        Large-scale dataset with diverse driving scenarios
        Contains images, videos, and GPS/IMU data
        Annotations for object detection, segmentation, and lane marking
    Cons:
        No LiDAR or RADAR data
        Some annotations might be noisy or incomplete

Waymo Open Dataset
    Task: Autonomous driving, object detection, and tracking
    Pros:
        High-quality and diverse dataset
        Multi-modal data with camera, LiDAR, and GPS/IMU
        Pre-processed sensor data for easier usage
    Cons:
        Large size requiring significant storage space
        Limited to Waymo vehicles and their specific sensors

ApolloScape
    Task: Autonomous driving, semantic segmentation, and instance segmentation
    Pros:
        Large-scale dataset with diverse driving scenarios
        High-resolution images and videos
        Detailed annotations for object detection and segmentation
    Cons:
        No LiDAR or RADAR data
        Complex dataset structure

---

YOLO (You Only Look Once)
    Task: Object detection
    Pros:
        Real-time object detection with high accuracy
        Single-pass architecture for faster processing
    Cons:
        Struggles with small object detection

Faster R-CNN
    Task: Object detection
    Pros:
        High accuracy in object detection
        Good for detecting small objects
    Cons:
        Slower than YOLO, not suitable for real-time applications

EfficientDet
    Task: Object detection
    Pros:
        High accuracy and efficient architecture
        Scalable with different model sizes
    Cons:
        Slower than YOLO but faster than Faster R-CNN

SegNet
    Task: Semantic segmentation
    Pros:
        Good for scene understanding tasks
        Efficient architecture with fast inference
    Cons:
        Less accurate than some other semantic segmentation models

DeepLab
    Task: Semantic segmentation
    Pros:
        High accuracy in semantic segmentation tasks
        Incorporates multi-scale features for better performance
    Cons:
        Slower inference than some other semantic segmentation models

Mask R-CNN
    Task: Instance segmentation
    Pros:
        High accuracy in object detection and segmentation
        Flexible architecture for various tasks
    Cons:
        Slower than other object detection models like YOLO

---

Object Detection: YOLOv5 or EfficientDet
Semantic Segmentation: DeepLabv3+ or SegNet
Object Tracking: DeepSORT or SORT

BDD100k: Good for object detection, lane markings, and traffic infrastructure, but it contains data from the United States.
ApolloScape: High-resolution images with semantic segmentation annotations, but data is from China.
a2d2 : High-resolution data from Germany, suitable for detecting traffic infrastructure elements, but has limited samples.

[saed13]

https://github.com/FabianGermany/AutonomousDrivingDetectron2

[saed13]

https://github.com/facebookresearch/detectron2