Migration away from DLIB

[Halmoni100]

@boltgolt I have a vested interest in getting the CNN feature to work on a cross-GPU/NPU-platform. Perhaps CPU SIMD for preprocessing and PyTorch/ONNX for the DL component? I'm a C++/signal-processing/ML developer and would be willing to do the work for that.

Edit: actually DLIB is probably still good for CPU stuff

[boltgolt]

I'm open to it, dlib was just the most available thing when i started Howdy 8 years ago at this point. I have not followed the space as closely but surely something better has come out since

[Halmoni100]

With regards to improving the DL approach to facial recognition, the source code for state-of-the-art tends to be proprietary, but the methods are published, so the challenge is to re-implement them in an open source code base.

It looks like in open-source land dlib's offerings are still considered state-of-the-art.

I think the current available toolchains' support for Linux and hardware will improve over time, but an better open-source facial recognition algorithm will have to be a deliberate open source effort, so that's probably where the effort should go for now. Preferably a full training to deployment pipeline.