OpenPose
Real-time multi-person pose estimation by CMU for body, hand, and face keypoints.
About
OpenPose came out of Carnegie Mellon University's Perceptual Computing Lab as the first real-time system to jointly detect body, foot, hand, and face keypoints for multiple people in a single image, up to 135 keypoints in total: 25 for body and feet, 21 per hand, and 70 for the face. It processes images, video files, webcams, and Flir industrial cameras, outputs overlays plus JSON, XML, or YML keypoint files, and exposes C++ and Python APIs alongside a command-line demo. Supported platforms include Ubuntu, Windows, macOS, and NVIDIA Jetson TX2, with CUDA, OpenCL, and CPU-only backends. Licensing is the main caveat: OpenPose is free only for non-commercial use, and commercial deployment requires a paid license through CMU's FlintBox portal. Despite newer alternatives, it remains widely used in research, motion analysis, and as a conditioning signal for pose-guided image generation.
Should you use OpenPose?
Pick it when
Pick OpenPose for non-commercial research, motion analysis, or art projects that need body, foot, hand, and face keypoints for several people from one system, or OpenPose-style skeletons as conditioning for pose-guided image generation.
Look elsewhere when
Skip it for any commercial product, since that requires a paid license from CMU. RTMPose under Apache 2.0 covers real-time whole-body keypoints, and MediaPipe is far easier for body and hand tracking on phones or in the browser.
Alternatives to OpenPose
- RTMPose
Apache 2.0 with 133-point whole-body models and MMDeploy export, so it can ship commercially where OpenPose needs a paid CMU license.
- DWPose
Apache 2.0 whole-body model on the RTMPose architecture with ONNX exports, a common OpenPose substitute for ControlNet pose conditioning without the non-commercial limit.
- MediaPipe
Runs on Android, iOS, and the web with no GPU required and is much simpler to set up, under Apache 2.0 instead of OpenPose's non-commercial terms.
- Ultralytics YOLO
Trains detection, tracking, and pose in one package under AGPL-3.0 rather than a non-commercial license, but its pretrained pose models are body-only, without face and hand points.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- Custom
- Minimum VRAM
- 6 GB
- Added
- Apr 3, 2026