Reproduce “Visible Feature Guidance for Crowd Pedestrian 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
Zhida Huang, Kaiyu Yue, Jiangfan Deng, Feng Zhou
The work from Aibee Inc. introduces Visible Feature Guidance (VFG) for pedestrian detection, using visible features to guide bounding box regression and a specialized NMS approach. This led to a 3.5% mAP increase and 4.5% recall improvement for Faster R-CNN on the Crowdhuman dataset, particularly enhancing detection in dense, occluded crowds.
Key insights
- Features derived from only the visible portion of an object provide a less ambiguous signal, improving the accuracy of both visible and full bounding box regression, especially under occlusion.
- Applying NMS to visible bounding boxes effectively differentiates distinct instances in dense crowds where full bounding boxes would heavily overlap, thereby preserving more correct detections.
- The Hungarian algorithm offers a robust and optimal method for associating semantic parts with their respective main instances, outperforming simpler greedy approaches for complex scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.