OpenPose Models: BODY_25 vs COCO, Output Format and Licence

This page covers what CMU's OpenPose repository says about its models and output, read on October 17, 2026. The licence is read in full in pose estimation model licenses, and the multi-person alternatives are in multi-person pose tracking.

Body models#

OpenPose's output document shows how to read the body-part order of any model from its C++ API, naming BODY_25, COCO_18, MPI_15, BODY_25B and BODY_135. Two have output-format diagrams in that document, and their orders, read from poseParameters.cpp, are:

• BODY_25: 25 body parts, from 0 Nose to 24 RHeel. See the BODY_25 keypoint map.
• COCO (PoseModel::COCO_18): 18 body parts, from 0 Nose to 17 LEar. See the OpenPose COCO 18 keypoint map.

BODY_25 contains every COCO 18 name plus MidHip and six foot points (LBigToe, LSmallToe, LHeel, RBigToe, RSmallToe, RHeel), and the shared names sit at different indices after index 7: RHip is 9 in BODY_25 and 8 in COCO 18. Each map ends with a Background entry, which the source excludes from the part count. OpenPose's COCO 18 is not the 17-keypoint COCO order MoveNet and YOLO pose use: it adds Neck at index 1 and spells sides as R and L, so the crosswalk does not join its points to COCO 17's.

The README's headline figure is 135 keypoints in total for body, hand, face and foot together, and the C++ API names a BODY_135 model.

Output format#

With --write_json, OpenPose writes one JSON file per frame with a people array. Each person's pose_keypoints_2d holds body part locations and detection confidence "formatted as x0,y0,c0,x1,y1,c1,...", in the model's body-part order. The --keypoint_scale flag sets the coordinate range, for example 0 to 1, -1 to 1, or source or output pixels, and the confidence runs from 0 to 1. Face and hand keypoints come in matching face_keypoints_2d, hand_left_keypoints_2d and hand_right_keypoints_2d arrays, and the 3D fields stay empty unless --3d is enabled.

Platforms#

The README lists Ubuntu 20, 18, 16 and 14, Windows 10 and 8, Mac OSX and the Nvidia TX2, with CUDA (Nvidia GPU), OpenCL (AMD GPU) and CPU-only builds, a command-line demo, C++ and Python APIs, and a portable Windows demo. It documents single-person 3D keypoint detection by triangulation from multiple views with Flir cameras, and says body runtime is invariant to the number of people detected while hand and face runtime depends on it.

Licence#

The LICENSE is headed "ACADEMIC OR NON-PROFIT ORGANIZATION NONCOMMERCIAL RESEARCH USE ONLY" and grants a license "to use the Software for noncommercial research purposes". The README says OpenPose "is freely available for free non-commercial use" and points commercial users to a FlintBox link. In our reading, a commercial fitness app needs that separate licence; the details are in pose estimation model licenses.

When it fits#

Our reading: OpenPose fits research and prototypes that want body, hands and face from one system, or that must emit BODY_25 for tools that expect it. For a commercial product the licence usually decides against it before the engineering does; OpenPose alternatives you can ship commercially covers the options, and multi-person pose tracking the architecture.