tfaehse
commented
2 years ago That's regrettable. I've just tested this, for now just a few notes:
- my input data is fairly bad, since it's just my dashcam footage. Faces are rare, especially high-def profile shots like these. This will hopefully get better, especially also because of a really nice data donation I received.
- The
_rectweights files don't perform well. I originally thought the higher resolution of training fixed-size images would benefit detection (especially of small objects, i.e. farther away plates), but Yolov5's mosaic data loader makes such a massive (positive) difference that it absolutely isn't the case.
For now, using _mosaic weights should yield better results. Long-term I'll hopefully be able to publish much better trained networks. I'll have a look at your PR and the other issue soon (I promise!), I'm unfortunately fairly busy at the moment.
edit: if you're looking to blur clips and only care about faces, that's a much simpler problem. There are some really nice, labelled datasets for facial recognition, in theory you can just train those with YOLOv5, name the face class "face" and drop them into the weights folder.
joshinils
i screen recorded a clip starting from https://youtu.be/EfdmAuB7JzU?t=296 and cut it down to one second. i have an amd gpu, so cuda does not work for me, cpu time for that one second clip was 3minutes 8sec.
none of the faces are blurred in the output clip. i used the 1080p_medium_rect weights and inference_size 1080 roi_multi of 1 quality of 10 threshold of 1 blur_size of 10 frame_memory of 1