OpenPose’s backroom: experimental training and whole-body models

What it does This repository holds the training scripts and validation tools used to build OpenPose models, plus a few experimental networks that never made it into the main release. It covers the full pipeline: generating training scripts, running the actual training, and evaluating accuracy. The authors warn that you should already know your way around the main OpenPose repo before stepping in here.

The interesting bit The real draw is the BODY_135 whole-body model, which jointly detects body, hand, face, and foot keypoints—135 in total—from a single network. There is also a BODY_25B variant that trades speed for accuracy compared to the standard OpenPose model. Both are explicitly tagged as experimental and may misbehave compared to the official releases.

Key highlights

• Ships training, testing, and validation scripts for OpenPose models.
• Includes the BODY_135 whole-body pose estimator (from the ICCV 2019 paper) and the higher-accuracy BODY_25B alternative.
• Licensed for non-commercial use only; commercial use requires a separate agreement through CMU’s FlintBox program.
• Originally tested on Ubuntu 16 with CUDA 8; the authors note it compiles on Ubuntu 20 with WSL2 but may need tweaks for newer CUDA versions.

Caveats

• The entire repository is explicitly labeled “highly experimental and not production ready.”
• Hardware support is dated: it was tested up to CUDA 10, so modern GPUs and newer Ubuntu versions may require modifications.
• The non-commercial license means you cannot legally use the resulting models in a commercial product without negotiating a separate license.

Verdict Researchers and academics who want to retrain OpenPose or experiment with whole-body keypoint detection should look here—if they are comfortable with legacy CUDA and an explicitly unstable codebase. Commercial teams and developers looking for a plug-and-play solution should stick to the main OpenPose release.