What it does

This repo implements two MPI-authored computer vision papers—DeeperCut (ECCV 2016) and ArtTrack (CVPR 2017)—for detecting human body keypoints in images. It handles both single-person and multi-person scenarios, with pre-trained models available for MPII and COCO datasets.

The interesting bit

The multi-person demo requires compiling dependencies first, which suggests there’s some Cython or native code bridging the gap between TensorFlow and the graph-based optimization tricks from the papers. The single-person path is simpler: download weights, run a script, hope your CUDA environment matches 2017-era expectations.

Key highlights

• Direct implementation of two well-cited academic papers (DeeperCut and ArtTrack)
• Pre-trained models for MPII (single-person) and COCO (multi-person) datasets
• Python 3.6 + TensorFlow 1.x stack, with explicit GPU selection via CUDA_VISIBLE_DEVICES
• Requires disabling cuDNN autotuning (TF_CUDNN_USE_AUTOTUNE=0) due to a known TensorFlow issue
• Training pipeline documented separately in models/README.md

Caveats

• Dependencies lock you to Python 3.6 and tensorflow-gpu (TensorFlow 1.x era, now deprecated)
• The README doesn’t specify which TensorFlow version, so expect archaeology
• Multi-person demo needs a compile step whose contents aren’t shown in the README

Verdict

Worth a look if you’re reproducing the DeeperCut/ArtTrack papers or need a baseline for multi-person pose estimation research. Skip it if you want a modern, maintained solution—this is a research artifact from the TensorFlow 1.x era, not a living project.