Openpose with TensorRT

[nguyenhoangthuan99]

I have successfully converted open pose 25 pose-model (version 1.6.0) to TensorRT but the issue is when I run inference in python, the post-process is very slow. I used post-processing code from here https://github.com/michalfaber/keras_Realtime_Multi-Person_Pose_Estimation. The net time is 0.2s/image with input size(368,496) but the post-process time is ~ 1s/image and the result is slightly worse than the original C++ version.

So I want to integrate my TensorRT model and use post-processing C++ code in the official version. Can you give me some advice or instruction?

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[fbens47]

Hi, @nguyenhoangthuan99, I'd like to try OpenPose with TensorRT too, but I am unable to find the output names of the network. Netron doesn't help, where did you get it?

Thanks

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[nireshd]

Hi, @nguyenhoangthuan99, I'd like to try OpenPose with TensorRT too, but I am unable to find the output names of the network. Netron doesn't help, where did you get it?

Thanks

hey @fbens47, did you end up finding the output names of the network?

[nguyenhoangthuan99]

Dear @fbens47, It’s been a long time and I don’t remember the details. But as I remember, I converted the model to onnx and use ONNx parser on python code, if you want to get the inputs and outputs name you can try to check this issue https://github.com/onnx/onnx/issues/2657.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[JJLimmm]

@nguyenhoangthuan99 @nireshd @fbens47 how did you guys replace the original pre-processing steps to your custom one (for feeding the input into the model) and how to post-process the model outputs?

Original Preprocessing steps using the python OP wrapper: frame = cv2.imread(file) datum.cvInputData = frame opWrapper.emplaceAndPop(op.VectorDatum([datum]))