ViTPose

Why we included this project

ViTPose is the official implementation of the NeurIPS 2022 paper on vision transformer baselines for human pose estimation, and the repo goes beyond the paper: it ships pretrained weights for several model sizes, training and inference configs, and integrates with the MMPose ecosystem, so you can move from the publication to a working keypoint detector without rebuilding everything by hand. The follow-up ViTPose++ extends the same backbone to generic body pose across multiple benchmarks, which helps if you want one approach that transfers across datasets rather than a patchwork of per-dataset models. With 81.1 AP on the MS COCO keypoint test-dev set and Hugging Face demos for both images and video, it's a solid baseline for research comparisons and a practical foundation for building keypoint detection into an application.