Dear Forum,
Has anyone tried to convert and run the deeptag-pytorch on the camera ? Do you think it is possible / worthwhile the effort ?
Thanks,
Tamás
Dear Forum,
Has anyone tried to convert and run the deeptag-pytorch on the camera ? Do you think it is possible / worthwhile the effort ?
Thanks,
Tamás
Hi @TamsLaszip
Not sure, but taking a quick look at the project, it looks like a two stage model, which should be convertible (if model layers are supported on VPU) to depthai blobs. The issue is you would need custom decoding for both models which will add additional overhead and will slow down the inference quite a bit.
Hi jakaskerl
The second stage does the decoding for me but the performance is my issue. I would need at least 5 FPS to use it in our application.
TamsLaszip
Well the first stage should be no problem, but you would need to do the decoding (second stage) on host then.
Hi jakaskerl
I have tested the process and it takes between 0.1 to 0.3 sec for the first stage, 0.1 for the second (decoding) on my laptop using a GPU, double the time with CPU only.
Based on the paper, it seems they use a MobileNet backbone which should be rather efficient on OAKs. If you want to give it a try, you can check their inference code and see where they initialize PyTorch models. Then you can use torch.onnx.export to export them to ONNX, and then use our blobconverter to convert to blob. I wouldn't focus on any accuracy, just to give the conversion a try. You can then benchmark this on OAK to see if it's suitable.
Feel free to join our Innovation Lab to stay up to date with upcoming products which will have more capabilities for AI inference.