Reproduce “RRTrack: Robust and Recoverable Object 6D Pose Tracking for Dynamic Scenes” 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

Junyue Li, Ye Zheng, Yifan Chen, Zhe Sun, Xuelong Li

RRTrack presents a unified framework for robust and recoverable 6D object pose tracking in dynamic and occluded scenes. The system improves pose accuracy and autonomously recovers from object disappearance, achieving an average 55.2 FPS on a new synthetic benchmark and high success rates on challenging real-world tasks without object-specific training.

Key insights

  • Tightly coupling 2D visual observations with 3D geometric consistency checks, through rendered-mask agreement, is crucial for preventing drift and maintaining robustness in dynamic environments.
  • A dual-bank feature retrieval system, combining diverse offline synthetic views with real-time online observations via foundation models, allows for efficient and training-free recovery from full object loss.
  • Adaptive thresholding, continuously updated from stable tracking data, provides a robust and responsive mechanism for managing tracker states and memory, improving recovery responsiveness.

Summary by alphaXiv.

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