pytorch-cpn

Why we included this project

If you work on pose estimation and want a codebase you can finish reading in an afternoon, this PyTorch port of the Cascaded Pyramid Network is a good find. The original model won the MSCOCO keypoints challenge in 2017, and its design stays easy to follow here: a GlobalNet handles the easy joints while a RefineNet circles back on the ones that are hard to locate, with the two-stage idea kept visible instead of buried under framework boilerplate. The repo includes COCO training and evaluation flows with per-backbone AP results for ResNet-50 and ResNet-101, so you can verify a reproduction before you trust it. For people getting into multi-person keypoint estimation, that combination of a small, readable codebase and published numbers beats starting from a sprawling production framework.