OpenPose Alternatives, Matched to the Reason You Are Leaving

Disclosure: this site is funded by KinesteX, which sells a commercial motion-tracking SDK; it does not appear in the list below, and none of the models here is ours. Every fact comes from each project's own repository, model card or documentation, read on October 22, 2026. What OpenPose itself outputs is covered in OpenPose models.

Why teams look for an OpenPose alternative#

• The licence. OpenPose's LICENSE is headed "ACADEMIC OR NON-PROFIT ORGANIZATION NONCOMMERCIAL RESEARCH USE ONLY" and grants a licence "to use the Software for noncommercial research purposes". It also says that derivatives and modifications you make "will be owned by Licensor". The README sends anyone wanting a commercial licence to a CMU FlintBox page. A paid app needs that separate deal or a different model.
• Its age. The last release in OpenPose's release notes is 1.7.0, dated November 15, 2020; the section after it is headed "Current version - Future OpenPose 1.7.1".
• Where it runs. The README lists Ubuntu, Windows, Mac OSX and the Nvidia TX2, with CUDA, OpenCL and CPU-only builds. It lists no Android or iOS build, so a phone app needs another model.

What OpenPose still does well#

Be fair to it before leaving. The README says body runtime is "invariant to number of detected people", it detects body, 2x21 hand and 70 face keypoints from one system, it offers BODY_25 with six foot points, and it documents single-person 3D by triangulating multiple camera views. For research, or for a pipeline built around BODY_25, it can still be the right tool.

The alternatives, by reason#

Several people under a licence you can ship: RTMPose#

RTMPose is the real-time project in OpenMMLab's MMPose, whose LICENSE is the Apache License, Version 2.0. Its README offers body models in a 17-keypoint COCO set, a 26-keypoint Halpe26 set and a 133-keypoint COCO-WholeBody set, and its pipeline table pairs each pose model with an RTMDet person detector. It deploys through ONNX Runtime, TensorRT, ncnn and OpenVINO, with Android examples on ncnn, and the rtmlib package "does not require Pytorch or MMCV". MMPose also publishes RTMO, which its README calls "a state-of-the-art real-time method for multi-person pose estimation". We found no separate terms for the pretrained checkpoints, so check those before shipping. See RTMPose models.

Several people, if AGPL-3.0 works for you: Ultralytics YOLO26 pose#

Ultralytics' default pose models predict the 17 COCO keypoints for every person detected, each with a box, in five sizes from n to x, and its export table covers CoreML and LiteRT for phones. The code is AGPL-3.0; the README sells an Enterprise License for products that cannot meet it. See Ultralytics YOLO pose models and MediaPipe vs YOLO pose.

Up to six people in a browser: MoveNet MultiPose#

MoveNet's TF.js README describes MULTIPOSE_LIGHTNING as a "Multi-pose detector that detects up to 6 poses", with optional tracking IDs across frames. The tfjs-models code is Apache-2.0, but the text of the MultiPose model card carries no licence line, so confirm its terms. See MoveNet models.

One person on a phone or in a browser: MediaPipe, MoveNet, ML Kit, Apple Vision#

• MediaPipe Pose Landmarker returns 33 landmarks with 3D world coordinates, on Android, iOS, web and Python, under Apache-2.0; its model card scopes it to one person.
• MoveNet SinglePose returns the 17 COCO keypoints, Lightning for speed and Thunder for accuracy, with an Apache-2.0 SinglePose card. MediaPipe vs MoveNet decides between the two.
• ML Kit Pose Detection returns 33 landmarks on Android and iOS, but Google's guide says it "can only detect one person in an image" and that the API "is offered in beta, and is not subject to any SLA or deprecation policy". See ML Kit pose detection.
• Apple Vision runs body pose with no model file to ship: Apple's documentation gives iOS 14 and macOS 11 as the first versions of its 2D body pose request. See Apple Vision body pose.

Not an escape: AlphaPose#

AlphaPose is one of the systems OpenPose's README compares its runtime against, and it supports COCO 17, Halpe 26 and Halpe 136 keypoint sets with tracking. Its licence is the same kind as OpenPose's: the LICENSE is headed "ACADEMIC OR NON-PROFIT ORGANIZATION NONCOMMERCIAL RESEARCH USE ONLY", with Shanghai Jiao Tong University as licensor, and the README says it is "freely available for free non-commercial use". Swapping OpenPose for AlphaPose does not fix a commercial licensing problem.

If you depend on BODY_25#

BODY_25 is OpenPose's 25-point body model: every point of its 18-point COCO model, Neck included, plus MidHip and six foot points (LBigToe, LSmallToe, LHeel and their right-side counterparts). No alternative here emits it as-is, so compare the points you use:

In our reading, Halpe26 has a counterpart for every BODY_25 point, which makes RTMPose's 26-keypoint models the closest commercial-friendly match; MMPose's Halpe26 config also adds a head point. Names and order differ throughout, so remap by hand rather than by index. Every index is in the BODY_25 keypoint map and the Halpe26 keypoint map.

At a glance#

Which should you pick?#

• A commercial product with several people in frame → RTMPose for an Apache-2.0 codebase, or YOLO26 pose with the AGPL-3.0 question settled.
• A pipeline that consumes BODY_25 foot points → RTMPose's Halpe26 models, remapped by name.
• One user on a phone → MediaPipe Pose Landmarker or MoveNet; native iOS apps can also use Apple Vision.
• Research only, and you need body, hands and face from one system → OpenPose's licence covers you, or RTMPose's 133-point wholebody models.

Each licence is read in full in pose estimation model licenses, and the tracking layer every multi-person option still needs is in multi-person pose tracking.