Reproduce “Detection in Crowded Scenes: One Proposal, Multiple Predictions” 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
Xuangeng Chu, Anlin Zheng, Xiangyu Zhang, Jian Sun
Researchers from Peking University and MEGVII Technology introduce a "Multiple Instance Prediction" (MIP) paradigm for object detection, enabling each proposal to predict a set of correlated instances rather than a single one, which improves detection in crowded scenes. This approach, incorporating an Earth Mover's Distance (EMD) Loss and Set NMS, boosts Averaged Precision by 4.9% on the CrowdHuman dataset and reduces the log-average Miss Rate by 1.0% on CityPersons.
Key insights
- The fundamental challenge in crowded scene detection arises from the 'one proposal, one prediction' paradigm, which forces models to differentiate highly similar features for individual, overlapping objects.
- Allowing proposals to predict a *set* of instances simplifies the learning process, as nearby proposals can then be trained to infer the same set of objects, intrinsically handling high overlap.
- Standard NMS is a major bottleneck in crowded scenarios, necessitating a modified NMS (Set NMS) to preserve multiple valid detections that originate from a single proposal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.