question / knowledge check on face identity detection

edit

Ok got it, dont mix things - thank you! In the method readNetFromTorch() theres is one comment i dont understand:

Also some equivalents of these classes from cunn, cudnn, and fbcunn may be successfully imported. Does this mean cuda and cudnn (have both in path) will be automatically used? If not i still can write a wrapper for torch too like i did it for yolo.

@holger, it's just a feature of Torch framework. You can serialize a model with CudaTensors and FloatTensors (like gpu and cpu mode). However you can not import model serialized in gpu mode with no CUDA installed (that means without Nvidia GPU). Hopefully there is no difference between them and OpenCV can import Torch models even if their owner forgot to switch them to cpu mode.

Got it, thank you for clarification.

So the cuda / cudnn optimization is in the model? I am no expert but this sounds like a bad design decision(maybe its really fast that way) - its a runtime aspect imho and should be transparent.

Sadly properietary cuda is a must have for me as it make a big big difference in performance. I need to be fast :-) https://www.youtube.com/watch?v=VTHsO...

Hello torch wrapper here i come. I find open cv really really awsome. Its a great source for having a reference implementation, way better than the examples coming with the cnn's.

Ok i read about the triplets loss https://omoindrot.github.io/triplet-loss

I also read about siamese networks https://www.quora.com/What-are-Siames...

@dkurt, @berak Now i understand the distance computing/ embedding you was talking about. I currently have the stupid idea to drop all this complexity(triplet mining & loss is not that easy for me) and just make a small model which overfits on persons(via transfer learning) Does this makes any sense and could this work?

Edit: Overfitting here would ultimativly mean that i can indentify only on same conditions(weather, environment). Bad - right?

Greetings + Thx again, Holg

Answering by myself (writing helps me thinking oo) - no - intentionally overfitting is a bad idea here. I will try out two things:

• Naive approach : try out and try transfer learning and see if it possible to introduce a new object class for one specific object. One shot learning is another term on which you can google.
• Serious approach: try to train a siamese network with triplets loss function for specific object class

Thank you for helping me coming to that conclusion.

Some peopel even use the feature vector (output from facenet) as input to train a model instead of computing distance directly. Well maybe a bit overkill but sounds consequent to me