The C++ Inference Library of OpenPose-Plus v2.0.0

[ganler]

Features

• New abstraction with better performance
  • Operator API: For users to manage detailed DNN inference and post-processing. (~90 FPS in video processing)
  • Stream API: Based on Operator API, we made a scheduler to do stream processing, provide end2end video processing. You can have it with very simple C++ stream operators like << or >> to set the input/output streams. (~120FPS in video processing).
• ONNX model parser support.
• Documentation.

The performance was tested on 1070 Ti with 6 physical core CPU.

WIP

• Documentation.
  • Doxygen Comments.
  • Doxygen to Sphinx via Breathe.
  • Deployment.
• Post-Processing.
  • Proposal Network(NMS).
• Benchmark over new models.
  • OpenPose model.
  • Lightweight OpenPose.
  • Proposal Network
• MISC.
  • Tests on ONNX parser.
  • Add serialized TensorRT Engine file.
  • Fix merge conflicts
  • CIs.
  • Ensure the scripts and docker running.
  • Provide a end2end simple docker application to generate a processed video.

Looking for preliminary reviews and contributions.

[ganler]

The documentation comments and conflicts are solved.
Do you have some time to look at the Travis CI? It seems to be a problem with CUDA.

[ganler]

@lgarithm I saw your comments about OpenCV on the CI scripts. Is there a problem when building OpenCV in docker? Is this helpful? https://hub.docker.com/r/schickling/opencv

[lgarithm]

The CI is failing because CUDA become a required dependency.
We used to have a stub (fake_openpose_plus.cpp) for testing purpose, which is now removed.

[lgarithm]

There are two options to fix:
1). use container with tensorRT and CUDA for CI.
2). revert fake_openpose_plus.cpp

[ganler]

@lgarithm I think for the best, we'd better select option 1. Because "fake_openpose_plus" may not simulate the real situation of compiling on GPU. Can we get this done by using nvidia-docker? (I'm not very familiar with docker. :(

[ganler]

I removed fake_openpose_plus because the older APIs are deprecated.

[ganler]

Thanks to @Gyx-One , we successfully converted the official OpenPose model to ONNX format which enables TensorRT execution. We firstly got 20 fps by inferring 20 batched images.

[lgarithm]

@ganler could you make files end with '\n'