infocom-tpo / PoseNet-CoreML

I checked the performance by running PoseNet on CoreML
MIT License
155 stars 35 forks source link

Implementation on windows #8

Open swatinair123 opened 5 years ago

swatinair123 commented 5 years ago

Hi , How can i convert this code to make it work on windows?

mbotsu commented 5 years ago

Hi. Running CoreML on Windows is a difficult story. In the case of Unity, it can run on Windows. Have a look at this. https://github.com/infocom-tpo/PoseNet-Unity

rwightman commented 5 years ago

Thanks to the conversion code in this repository, I started on a path that ended up in a Python Tensorflow and PyTorch port of TFJS PoseNet. They both work on Windows with a recent version of Anaconda (tested with 64-bit Anaconda 2018.12 Python 3.7 installer).

https://github.com/rwightman/posenet-python https://github.com/rwightman/posenet-pytorch

Raw throughput of the neural net + post processing is around 100fps for the Tensorflow version on Windows with a 1080 Ti and modest 4 core CPU. PyTorch version is really slow in Windows for some reason (~65fps), but faster than Tensorflow (130+ fps) in Linux.

The webcam demo with skeleton and keypoints overlay keeps up with a 30fps webcam on either Windows or Linux.