Winning pose estimation by keeping features local

What it does

RSN is a PyTorch implementation of the multi-person pose estimator that won the COCO 2019 Keypoint Challenge and the COCO 2019 Best Paper Award. It locates human keypoints by preserving fine-grained spatial details rather than compressing them through repeated downsampling, and adds a Pose Refine Machine—an attention module—to sharpen joint coordinates.

The interesting bit

Instead of the classic pyramid that mixes features across wildly different resolutions, RSN aggregates “intra-level” features at the same spatial size. The authors argue that delicate local representations—rich in low-level texture—are what let the model place elbows, wrists, and ankles exactly where they belong.

Key highlights

• Took 1st place in the COCO 2019 Human Keypoint Detection Challenge and won the COCO 2019 Best Paper Award.
• Scored 78.6 AP on COCO test-dev with a single model, and 79.2 AP with an ensemble; also hits 93.0 mean accuracy on MPII test.
• Achieved these results without extra training data or pretrained backbones.
• The architecture has been upstreamed into the OpenMMLab MMPose framework.
• The authors note their original experiments ran on an internal deep-learning platform, so this PyTorch reimplementation produces slightly different metrics.

Caveats

• The README explicitly warns that the PyTorch version’s numbers diverge from the paper because the original work used a custom DL platform.
• All models were trained on 8 V100 GPUs, so reproducing results demands serious hardware.

Verdict

A strong reference if you are researching keypoint architectures or spatial feature design; for production use, the authors themselves now direct users to the MMPose integration instead of this standalone repo.