RTMPose
Real-time multi-person pose estimation by OpenMMLab with high accuracy.
About
RTMPose is a family of real-time pose estimation models from OpenMMLab, distributed inside MMPose, a PyTorch toolbox covering 2D multi-person body pose, hand and facial keypoints, whole-body estimation with 133 keypoints, animal pose, and 3D mesh recovery. The models are engineered for practical deployment, balancing accuracy against latency across multiple sizes, and the line has grown to include RTMO for one-stage real-time multi-person estimation, RTMW whole-body models with flexible input resolutions, and RTMW3D for real-time 3D whole-body pose. MMPose supplies the surrounding infrastructure: a model zoo, training tools with FP16 support, standard benchmarks such as COCO and MPII, and export through MMDeploy to production runtimes. Everything is open source under the Apache 2.0 license. Researchers use the toolbox to develop new methods, while practitioners deploy RTMPose for motion capture, fitness, AR, and animation pipelines.
Should you use RTMPose?
Pick it when
Pick RTMPose when you need commercially usable real-time body, hand, face, or 133-point whole-body keypoints for motion capture, fitness, or AR, and are willing to work inside MMPose to train models and export through MMDeploy.
Look elsewhere when
Skip it if you want pose on a phone or in the browser with minimal setup, where MediaPipe is far simpler, or if the OpenMMLab config and dependency stack is more than a quick prototype needs.
Alternatives to RTMPose
- MediaPipe
Easiest route to pose and hand tracking on Android, iOS, and the web without a GPU, but it lacks animal pose and the training and benchmark tooling MMPose provides.
- OpenPose
Older joint body, hand, and face system with a familiar skeleton format, but non-commercial unless you buy a CMU license, where RTMPose is Apache 2.0.
- DWPose
Distilled whole-body models on the RTMPose architecture with ONNX exports that run without MMPose, narrower in scope but lighter to drop into generation pipelines.
- Ultralytics YOLO
Simpler to train detection and pose together in one CLI, but AGPL-3.0 where RTMPose is Apache 2.0, and it lacks MMPose's specialized whole-body and 3D models.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Intermediate (3/5)
- License
- Apache-2.0
- Minimum VRAM
- 4 GB
- Added
- Apr 3, 2026