RegTrack: Simplicity Beneath Complexity in Robust Multi-Modal 3D Multi-Object Tracking

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Existing 3D multi-object tracking (MOT) methods often sacrifice efficiency and generalizability for robustness.
• Complex association metrics and class-specific priors limit the performance of current 3D MOT techniques.

Purpose of the Study:

• To propose RegTrack, a robust, efficient, and generalizable method for multi-modal 3D MOT.
• To challenge the notion that increased complexity is essential for enhanced robustness in 3D MOT.

Main Methods:

• RegTrack formulates 3D MOT as motion-compensated representation learning, inspired by Yang-Mills gauge theory.
• It employs a unified tri-cue encoder (UTEnc) with point cloud and geometry encoders, guided by a pretrained vision-language model during training.
• Motion-induced variations in point-cloud representations are adaptively compensated using geometric cues.

Main Results:

• RegTrack achieves robust, efficient, and generalizable inference using only point-cloud inputs.
• The method boasts a small parameter count (2.67M) and outperforms 35 competing methods on KITTI and nuScenes datasets.
• Pairwise similarities of motion-compensated representations provide a simple yet effective association metric.

Conclusions:

• RegTrack presents a paradigm shift in 3D MOT by achieving high performance through efficient, motion-compensated representation learning.
• The method demonstrates superior robustness, efficiency, and generalizability compared to existing approaches.
• RegTrack's success highlights the potential of physics-inspired theories in advancing computer vision tasks.