Reproduce “Offline-Poly: A Polyhedral Framework For Offline 3D 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
Xiaoyu Li, Yitao Wu, Xian Wu, Haolin Zhuo, Lijun Zhao, Lining Sun
Offline-Poly presents a learning-free, polyhedral framework for offline 3D multi-object tracking, leveraging a Tracking-by-Tracking (TBT) paradigm to process arbitrary upstream tracker outputs. The method achieves state-of-the-art performance on the nuScenes (77.6% AMOTA) and KITTI (83.00% HOTA) datasets by systematically exploiting the resource unconstrainedness and future observability inherent in offline settings.
Key insights
- Offline tracking can be fundamentally improved by systematically and explicitly exploiting its intrinsic properties: 'resource unconstrainedness' (allowing computationally intensive global reasoning) and 'future observability' (enabling trajectory reasoning over the full temporal horizon).
- A 'Tracking-by-Tracking' (TBT) paradigm provides a robust and generalizable design, decoupling the offline tracker from specific upstream detectors or online trackers, thereby enhancing modularity and compatibility.
- Hierarchical matching and fusion mechanisms, including re-identifying fragmented tracklets (STWO), disentangling erroneous merges (STW), and integrating multiple tracker outputs (MTM), along with multi-perspective trajectory refinement, are crucial for achieving high-quality 4D auto-labeling.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.