Volodymyr

using depthData it should be trivial to see the difference. A picture is a flat object, while a real face is not.

because InPlaceABNSync is a custom operation implemented in CUDA - it cannot be exported to ONNX. To export the model, however, you can perform conv-bn fusion if your architecture allows...

On RTX3090 in float32 it runs at around 57 imgs/sec during inference.

Same issue on my system - CPU usage is close to 100%, GPU usage close to 0

Hi @AmitMY, sorry for the long wait, this project got kinda low on the priority list. Are you still interested in improving the pipeline? If so - please send me...

you're on master branch, right? The code doesn't seem to have any recursive calls. Can you run a debugger and set a breakpoint in sigm, step outside of it and...

better clip the value under `exp`

so?

Sure, you can give it a go.