/openpose

OpenPose: Real-time multi-person keypoint detection library for body, face, hands, and foot estimation

Primary LanguageC++OtherNOASSERTION


Default Config CUDA (+Python) CPU (+Python) OpenCL (+Python) Debug Unity
Linux Status Status Status Status Status Status
MacOS Status Status Status Status Status
Windows Status

OpenPose represents the first real-time multi-person system to jointly detect human body, hand, facial, and foot keypoints (in total 135 keypoints) on single images.

It is authored by Gines Hidalgo, Zhe Cao, Tomas Simon, Shih-En Wei, Hanbyul Joo, and Yaser Sheikh. Currently, it is being maintained by Gines Hidalgo and Yaadhav Raaj. In addition, OpenPose would not be possible without the CMU Panoptic Studio dataset. We would also like to thank all the people who helped OpenPose in any way. The main contributors are listed in doc/contributors.md.


Authors Gines Hidalgo (left) and Hanbyul Joo (right) in front of the CMU Panoptic Studio

Features

  • Functionality:
    • 2D real-time multi-person keypoint detection:
      • 15 or 18 or 25-keypoint body/foot keypoint estimation. Running time invariant to number of detected people.
      • 6-keypoint foot keypoint estimation. Integrated together with the 25-keypoint body/foot keypoint detector.
      • 2x21-keypoint hand keypoint estimation. Currently, running time depends on number of detected people.
      • 70-keypoint face keypoint estimation. Currently, running time depends on number of detected people.
    • 3D real-time single-person keypoint detection:
      • 3-D triangulation from multiple single views.
      • Synchronization of Flir cameras handled.
      • Compatible with Flir/Point Grey cameras, but provided C++ demos to add your custom input.
    • Calibration toolbox:
      • Easy estimation of distortion, intrinsic, and extrinsic camera parameters.
    • Single-person tracking for further speed up or visual smoothing.
  • Input: Image, video, webcam, Flir/Point Grey and IP camera. Included C++ demos to add your custom input.
  • Output: Basic image + keypoint display/saving (PNG, JPG, AVI, ...), keypoint saving (JSON, XML, YML, ...), and/or keypoints as array class.
  • OS: Ubuntu (14, 16), Windows (8, 10), Mac OSX, Nvidia TX2.
  • Training and datasets:
  • Others:
    • Available: command-line demo, C++ wrapper, and C++ API.
    • Python API.
    • Unity Plugin.
    • CUDA (Nvidia GPU), OpenCL (AMD GPU), and CPU-only (no GPU) versions.

Latest Features

For further details, check all released features and release notes.

Results

Body and Foot Estimation


Testing the Crazy Uptown Funk flashmob in Sydney video sequence with OpenPose

3-D Reconstruction Module (Body, Foot, Face, and Hands)


Testing the 3D Reconstruction Module of OpenPose

Body, Foot, Face, and Hands Estimation


Authors Gines Hidalgo (left image) and Tomas Simon (right image) testing OpenPose

Unity Plugin


Tianyi Zhao and Gines Hidalgo testing their OpenPose Unity Plugin

Runtime Analysis

Inference time comparison between the 3 available pose estimation libraries: OpenPose, Alpha-Pose (fast Pytorch version), and Mask R-CNN:

This analysis was performed using the same images for each algorithm and a batch size of 1. Each analysis was repeated 1000 times and then averaged. This was all performed on a system with a Nvidia 1080 Ti and CUDA 8. Megvii (Face++) and MSRA GitHub repositories were excluded because they only provide pose estimation results given a cropped person. However, they suffer the same problem than Alpha-Pose and Mask R-CNN, their runtimes grow linearly with the number of people.

Contents

  1. Features
  2. Latest Features
  3. Results
  4. Installation, Reinstallation and Uninstallation
  5. Quick Start
  6. Output
  7. Speeding Up OpenPose and Benchmark
  8. Training Code and Foot Dataset
  9. Send Us Failure Cases and Feedback!
  10. Citation
  11. License

Installation, Reinstallation and Uninstallation

Windows portable version: Simply download and use the latest version from the Releases section.

Otherwise, check doc/installation.md for instructions on how to build OpenPose from source.

Quick Start

Most users do not need the OpenPose C++/Python API, but can simply use the OpenPose Demo:

  • OpenPose Demo: To easily process images/video/webcam and display/save the results. See doc/demo_overview.md. E.g., run OpenPose in a video with:
# Ubuntu
./build/examples/openpose/openpose.bin --video examples/media/video.avi
:: Windows - Portable Demo
bin\OpenPoseDemo.exe --video examples\media\video.avi

Output

Output (format, keypoint index ordering, etc.) in doc/output.md.

Speeding Up OpenPose and Benchmark

Check the OpenPose Benchmark as well as some hints to speed up and/or reduce the memory requirements for OpenPose on doc/speed_up_openpose.md.

Training Code and Foot Dataset

For training OpenPose, check github.com/CMU-Perceptual-Computing-Lab/openpose_train.

For the foot dataset, check the foot dataset website and new OpenPose paper for more information.

Send Us Failure Cases and Feedback!

Our library is open source for research purposes, and we want to continuously improve it! So please, let us know if...

  1. ... you find videos or images where OpenPose does not seems to work well. Feel free to send them to openposecmu@gmail.com (email only for failure cases!), we will use them to improve the quality of the algorithm!
  2. ... you find any bug (in functionality or speed).
  3. ... you added some functionality to some class or some new Worker subclass which we might potentially incorporate.
  4. ... you know how to speed up or improve any part of the library.
  5. ... you have a request about possible functionality.
  6. ... etc.

Just comment on GitHub or make a pull request and we will answer as soon as possible! Send us an email if you use the library to make a cool demo or YouTube video!

Citation

Please cite these papers in your publications if it helps your research. Most of OpenPose is based on [8765346]. In addition, the hand and face keypoint detectors are a combination of [8765346] and [Simon et al. 2017] (the face detector was trained using the same procedure than the hand detector).

@article{8765346,
  author = {Z. {Cao} and G. {Hidalgo Martinez} and T. {Simon} and S. {Wei} and Y. A. {Sheikh}},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  title = {OpenPose: Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields},
  year = {2019}
}

@inproceedings{simon2017hand,
  author = {Tomas Simon and Hanbyul Joo and Iain Matthews and Yaser Sheikh},
  booktitle = {CVPR},
  title = {Hand Keypoint Detection in Single Images using Multiview Bootstrapping},
  year = {2017}
}

@inproceedings{cao2017realtime,
  author = {Zhe Cao and Tomas Simon and Shih-En Wei and Yaser Sheikh},
  booktitle = {CVPR},
  title = {Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields},
  year = {2017}
}

@inproceedings{wei2016cpm,
  author = {Shih-En Wei and Varun Ramakrishna and Takeo Kanade and Yaser Sheikh},
  booktitle = {CVPR},
  title = {Convolutional pose machines},
  year = {2016}
}

Links to the papers:

License

OpenPose is freely available for free non-commercial use, and may be redistributed under these conditions. Please, see the license for further details. Interested in a commercial license? Check this FlintBox link. For commercial queries, use the Contact section from the FlintBox link and also send a copy of that message to Yaser Sheikh.