What it does Trains CPM and Hourglass pose-estimation models, then squeezes them through MobileNet V2 inverted residuals so they run on a Snapdragon 845 at ~60 FPS. Ships with Android (Mace GPU or TFLite CPU) and iOS (CoreML) demo apps, plus pre-converted frozen graphs, TFLite, and CoreML models.

The interesting bit The accuracy trade-off is explicit and measured: PCKh scores of 93.78 (CPM) and 91.81 (Hourglass) at only 0.5 GFLOPs. The author notes both architectures and dataset “still have a huge margin of improvement”—a rare admission that doubles as a roadmap.

Key highlights

• Training code with nvidia-docker support and TensorBoard logging
• Android GPU path via Xiaomi’s Mace framework (v0.9.0, macOS-only build host)
• iOS path via community-contributed CoreML wrapper
• Pre-built APKs for quick device testing
• 22,446 training examples from AI Challenger, converted to COCO format

Caveats

• Mace demo builds only on macOS; Windows is unsupported and environment-specific build errors are expected
• TensorFlow >= 1.4 dependency shows the repo’s age; TFLite conversion instructions reference TF 1.9
• Single-person only; multi-person scenes need not apply

Verdict Good starting point if you need a hackable mobile pose baseline and can tolerate 2018-era tooling. Skip if you want turnkey multi-person detection or modern TF Lite/ML Kit integration.