body pose model? about dlib-models

[arrufat]

I was also facing a similar problem, I tried to train a shape predictor for pose estimation, but I couldn't make it work either, especially for weird poses (it turns out some humans are really flexible)

Dlib has this loss layer: loss_mean_squared_per_channel_and_pixel, which is what they use in this approach to estimate landmark heatmaps: Simple Baselines for Human Pose Estimation and Tracking.
When used with a ResNet50 backbone and an input size of 224x224, you can easily get inference times below 10 ms on a GPU, however, I don't think that would be suitable for an Android app, where there's no CUDA...

[fm64hylian]

Thank for your reply, it is indeed difficult without cuda(I am using windows with an Intel chip which makes it even worse) I haven't dig too much in models but I tried to run some like the body25 from open pose, I am completely new to deep learning so I kind of need a lead, is it possible to create a rough body pose model within one week? what would you recommend? I know that unity supports nn files as well, should I dig on tensorflow maybe? (they gave me 3 weeks to make this app lol)

[arrufat]

I am afraid I am not in a position to give you advice on how you should do your work.
However, I will say that, given the amount of time you have left, you'd be better off using some already trained solutions like this one: https://github.com/edvardHua/PoseEstimationForMobile

[fm64hylian]

thanks so much! I will definetely give it a look. Another super noob question, the steps for training the model are there but, is it there like a generated file that can just drag to read it with opencv or similar? or do I have to compile something to generate that? I still don't understand how most pretrained models work