Joint 3D Instance Segmentation and Object Detection for Autonomous Driving

Joint 3D Instance Segmentation and Object Detection for Autonomous Driving
Brief introduction~~
Framework PointNet++ as backbone.

• Backbone segmented the semantic (FG/BG) information for each point after extracting features and predicted the information of BBox (offset, dim, θ).
• Offset and cluster the foreground points, assign an ID to each category of points as an instance, and obtain a BBox by averaging the information of the first 5 predictions for each instance.
• Instance-aware RoI polling: Foreground points that are not in BBox but have corresponding IDs are also used for Refinement; foreground points that are in BBox but do not have corresponding IDs are deleted.(That is, each object proposal in Refinement considers the point of the instance instead of the point in the Box.)