Reproduce “Explicit Box Detection Unifies End-to-End Multi-Person Pose Estimation” 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
Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, Lei Zhang
The ED-Pose framework unifies end-to-end multi-person pose estimation by treating both human and keypoint detection as explicit box detection tasks within a cascaded Transformer architecture. It achieves state-of-the-art performance on CrowdPose (76.6 AP) and COCO (71.6 AP with ResNet-50), becoming the first L1 regression-based, post-processing-free method to outperform traditional heatmap-based Top-Down approaches.
Key insights
- Explicit human detection, serving as an initialization stage, significantly accelerates training convergence and provides crucial global context for subsequent keypoint detection.
- Representing keypoints as 4D bounding boxes (x, y, w, h) rather than just 2D points is critical for capturing rich local contextual information, leading to a substantial performance boost.
- Carefully controlled interactive learning within the decoder, where external keypoint-keypoint interactions are masked, is essential for maintaining robustness and preventing performance degradation, especially in crowded environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.