Real-time pose tracker that doesn't choke on crowds

What it does AlphaPose detects and tracks human keypoints for multiple people in images or video. It outputs 2D skeletons ranging from 17 standard COCO joints up to 136 whole-body keypoints including fingers and facial features, and can lift these into 3D meshes via an integrated HybrIK module. A bundled tracker called PoseFlow matches identities across frames so you get consistent skeleton IDs over time.

The interesting bit The project bundles several distinct research contributions—CrowdPose for overlapping bodies, Halpe whole-body keypoints, and HybrIK 3D shape estimation—into a single inference pipeline with a model zoo and a high-level Python API. That makes it less a single model and more a maintained research platform that happens to run at 20–23 fps.

Key highlights

• Claims first open-source results above 70 mAP on COCO (73.3 AP in results table) and 80 mAP on MPII (82.1 AP in table)
• PoseFlow tracker claims first open-source online tracking above 60 mAP and 50 MOTA on PoseTrack
• Supports whole-body estimation (face, hands, feet) via 26 or 136 keypoint models
• Integrates HybrIK for 3D human pose and SMPL mesh reconstruction
• Runs on Linux and Windows, with Colab notebooks and YOLOX detector support

Caveats

• License is strictly non-commercial; commercial use requires contacting the authors for a separate agreement
• Several roadmap items remain unfinished, including a PyTorch C++ version and dense support
• Speed claims (20–23 fps) reference older releases and specific backbones; current performance depends on hardware and model choice

Verdict Worth a look if you need off-the-shelf 2D/3D pose estimation or video tracking for research and prototyping. Skip it if you need a drop-in commercial solution or a lightweight edge deployment.