A no-frills PyTorch gym for training pose estimators

What it does

PyTorch-Pose is a training and evaluation pipeline for 2D single-person human pose estimation. It bundles data loaders with augmentation for MPII, LSP, FLIC, and MSCOCO, plus scripts for training, testing, and computing PCKh@0.5 scores. You get multi-GPU support, logging, and visualization out of the box.

The interesting bit

The project is essentially a careful re-implementation and glue layer around the original Stacked Hourglass training code, ported to PyTorch with modern conveniences like multi-threaded data loading. It also includes a later addition: Xiao et al.’s “Simple Baselines” model from ECCV 2018, so you’re not locked into hourglass architectures.

Key highlights

• Supports Stacked Hourglass networks and Simple Baselines (Xiao et al. ECCV 2018)
• Datasets: MPII, LSP, FLIC, and single-person MSCOCO keypoints
• Pretrained 2-stack and 8-stack hourglass models available via Google Drive
• Evaluation in both MATLAB and Python (ported from Tompson et al. CVPR 2015)
• Training curves and network output visualization built in

Caveats

• Developed for Python 2; Python 3 compatibility is untested according to the README
• Requires manual dataset setup: symbolic links for MPII images, Google Drive downloads for annotations, and a separate COCO readme
• PyTorch compatibility note says 0.4.1/1.0, which may need verification on current PyTorch versions

Verdict

Worth a look if you need a reproducible, well-documented starting point for single-person pose estimation research or want to benchmark against established MPII baselines. Skip it if you need multi-person pose tracking or a turnkey Python 3 solution without manual dataset wrangling.