This repository provides an official PyTorch implementation of the HRNet architecture for human pose estimation, as detailed in the CVPR 2019 paper. It is designed for researchers and practitioners in computer vision focused on accurate and spatially precise human pose estimation. The key benefit is the network's ability to maintain high-resolution representations throughout the entire process, leading to improved keypoint detection accuracy.

How It Works

The HRNet architecture maintains high-resolution representations by starting with a high-resolution subnetwork and progressively adding parallel subnetworks that process information at different resolutions. Repeated multi-scale fusions allow information exchange across these parallel streams, enriching the high-resolution representations. This contrasts with typical approaches that downsample early and then attempt to recover high resolution.

Quick Start & Requirements

• Install: Clone the repo, pip install -r requirements.txt, cd lib && make. Install COCOAPI (git clone https://github.com/cocodataset/cocoapi.git && cd cocoapi/PythonAPI && make install).
• Prerequisites: PyTorch >= v1.0.0, Python 3.6, Ubuntu 16.04, NVIDIA GPUs (tested on 4x P100).
• Data: Download MPII and COCO datasets and organize them as specified. Pretrained models are available via Google Drive or OneDrive.
• Links: Official Docs, Model Zoo

Highlighted Details

• Achieves state-of-the-art results on MPII and COCO keypoint detection benchmarks.
• Offers multiple HRNet variants (w32, w48) and corresponding pretrained models.
• Supports training and testing pipelines for both MPII and COCO datasets.
• Includes visualization tools for prediction results.

Maintenance & Community

The project is associated with authors from Tsinghua University and has seen updates and extensions like HRNet-DEKR and HigherHRNet. It has been cited in multiple influential papers.

Licensing & Compatibility

The repository does not explicitly state a license in the README. However, it is common for academic research code to be under a permissive license like MIT or BSD, but this should be verified.

Limitations & Caveats

The code was developed and tested on Ubuntu 16.04 with specific NVIDIA GPUs (P100s); compatibility with other platforms or GPU architectures is not guaranteed. The README mentions that PyTorch versions < v1.0.0 require disabling cuDNN BatchNorm implementations.