Reproduce “OTA: Optimal Transport Assignment for Object Detection” 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
Zheng Ge, Songtao Liu, Zeming Li, Osamu Yoshie, Jian Sun
Optimal Transport Assignment (OTA) re-frames label assignment in object detection as a global optimal transport problem, addressing limitations of independent label assignment. This approach achieves improved object detection accuracy, particularly excelling in scenarios with dense or overlapping objects.
Key insights
- Reframing label assignment as a global Optimal Transport problem offers a principled way to resolve ambiguous anchor conflicts, outperforming local, heuristic rules.
- A dynamic estimation of 'k' (number of positive anchors per ground truth) significantly improves performance compared to fixed 'k' values, adapting better to diverse object characteristics.
- The Optimal Transport framework can effectively adapt to the one-to-many assignment paradigm of dense CNN-based detectors, unlike direct adaptations of one-to-one global methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.