Reproduce “V2F-Net: Explicit Decomposition of Occluded 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

Mingyang Shang, Dawei Xiang, Zhicheng Wang, Erjin Zhou

Researchers from MEGVII Technology and Texas A&M University developed V2F-Net, which explicitly decomposes occluded pedestrian detection into sequential visible region detection and full-body estimation. This approach yielded a 5.63% AP increase and 3.89% MR^-2 reduction on CrowdHuman compared to a Faster R-CNN baseline, improving detection accuracy in crowded scenes.

Key insights

  • Explicitly decomposing the occluded pedestrian detection problem into visible region detection followed by full-body estimation simplifies the learning task and yields more robust results.
  • Performing NMS on visible bounding boxes provides a more reliable metric for distinguishing individuals in crowded scenes, leading to fewer false suppressions compared to NMS on full-body boxes.
  • Incorporating part-aware supervision through an auxiliary module during training, which is discarded during inference, can effectively guide the network to infer occluded body parts without adding computational overhead at test time.

Summary by alphaXiv.

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