Weakly and Self-Supervised Class-Agnostic Motion Prediction for Autonomous Driving.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Autonomous driving requires accurate motion prediction in dynamic environments.
• Class-agnostic motion prediction from LiDAR point clouds is a key research area.
• Current methods often rely on extensive motion annotations.

Purpose of the Study:

• To investigate weakly and self-supervised class-agnostic motion prediction using LiDAR.
• To reduce the reliance on detailed motion annotations by leveraging scene structure.
• To develop robust methods balancing annotation effort and prediction performance.

Main Methods:

• Proposed a weakly supervised paradigm using foreground/background masks for motion prediction.
• Utilized non-ground/ground masks as a less annotation-intensive alternative.
• Developed a self-supervised method requiring no annotations.
• Introduced a Robust Consistency-aware Chamfer Distance loss for outlier suppression.

Main Results:

• Weakly and self-supervised models outperformed existing self-supervised methods.
• Weakly supervised models achieved performance comparable to some supervised methods.
• Demonstrated effective balance between annotation effort and predictive performance.

Conclusions:

• Leveraging scene parsing cues (foreground/background, non-ground/ground) enables effective weakly and self-supervised motion prediction.
• Reduced annotation requirements significantly improve the practicality of motion prediction models.
• The proposed methods offer a promising direction for efficient autonomous driving perception.