An Effective Motion-Centric Paradigm for 3D Single Object Tracking in Point Clouds.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Driving

Background:

• 3D single object tracking (LiDAR SOT) is vital for autonomous driving.
• Current Siamese-based trackers struggle with textureless and incomplete LiDAR data, hindering appearance matching.
• Existing methods often neglect crucial motion information.

Purpose of the Study:

• To introduce a new motion-centric paradigm for LiDAR SOT.
• To develop a matching-free, two-stage tracker (M²-Track) that leverages motion clues.
• To explore semi-supervised learning strategies for LiDAR SOT.

Main Methods:

• A two-stage tracking approach: motion transformation for localization and motion-assisted shape completion for refinement.
• Incorporation of pseudo-label-based motion augmentation and self-supervised loss for semi-supervised learning.
• Evaluation on KITTI, NuScenes, and Waymo Open Dataset.

Main Results:

• M²-Track significantly outperforms state-of-the-art methods in fully-supervised settings across three datasets, achieving precision gains of ~3%, ~11%, and ~22% respectively.
• The tracker operates efficiently at 57FPS.
• In semi-supervised settings, M²-Track matches or surpasses fully-supervised performance with less than half the labeled data on KITTI.

Conclusions:

• The motion-centric paradigm offers a promising alternative to appearance-based tracking for LiDAR SOT.
• M²-Track demonstrates strong generalizability, efficiency, and effectiveness in both supervised and semi-supervised learning scenarios.
• The approach shows potential for auto-labeling and unsupervised domain adaptation in autonomous driving.