simple-HRNet

Why we included this project

HRNet's official research code is powerful but not exactly friendly to pick up, and this project exists to fix that. It wraps the high-resolution network in a small SimpleHRNet class that loads the official pre-trained weights and returns joint predictions for a single image or a whole batch, then adds multi-person detection by pairing HRNet with YOLOv3, YOLOv3-tiny, or YOLOv5. Ready-to-run scripts cover a live webcam demo, pulling keypoints to CSV from recorded video, TensorRT conversion for faster inference, and training on COCO, and a Colab notebook walks through the whole flow. The code runs on Windows and Linux and produces results that match the original implementation, which makes it a comfortable base for prototyping pose-based tools like fitness tracking or ergonomics analysis. Just treat it as a starting point rather than a maintained production library, and plan to adapt it for your own deployment.