Replace squeezenet DNN with pose tracking DNN about winmlexperiments HOT 4 CLOSED

Hi @cwule! The computational resources on the HL1 & 2 are likely not enough to run a larger model on the device. I would suggest trying out a streaming protocol to offload some of the processing to a companion pc.

from winmlexperiments.

Thanks @doughtmw for the tip. I also looked at the streaming option and Microsoft Research Cambridge showed an example where they used the streaming option for pose tracking on the Hololens 1. Unfortunately I need an offline solution. Since there are some smaller pose estimation networks available as well that run on mobile devices, I want to try to get it to run on the Hololens (even at a low framerate).

from winmlexperiments.

Sorry. Just saw that question. Sounds doable but also agreeing that it might be too heavy for the HoloLens. On the other hand if you have a mobile one it might work.

The TextAsset loading is really just a workaround for Unity's weird behavior. I don't recall exactly why but IIRC there was comment in the code I added that explains the reasoning.
It's likely possible with a newer Unity version to load the model the right way. Even better would be to load directly with the UWP stream IO APIs.

from winmlexperiments.

@reneschulte Thanks for the tip, I will look into the UWP IO APIs. The OpenCV DNN module would also be comfortable to use, but that doesn't work with UWP + ARM64 unfortunately.

from winmlexperiments.