Reproduce “CAMOT: Camera Angle-aware Multi-Object Tracking” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Felix Limanta, Kuniaki Uto, Koichi Shinoda

CAMOT introduces a camera angle-aware multi-object tracking method that estimates camera elevation angles and per-object depths solely from 2D bounding box detections. This lightweight approach enables pseudo-3D tracking and angle-aware association, achieving state-of-the-art performance with 63.8% HOTA on MOT17 and 62.8% HOTA on MOT20, while maintaining near real-time inference speeds of 27.9 FPS.

Key insights

  • Accurate camera elevation angles and object depths can be efficiently inferred from standard 2D bounding box detections alone, enabling effective pseudo-3D tracking without relying on deep learning depth estimators or specialized hardware.
  • Integrating camera angle and depth information significantly improves MOT performance by better handling occlusion and enhancing association accuracy, particularly reducing identity switches in crowded scenes.
  • A task-specific, geometry-based approach like CAMOT can be more robust and computationally efficient than general-purpose deep learning depth estimators for specific scenarios, such as high-angle surveillance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.