Closed lxy5513 closed 3 years ago
Hi, unfortunately, no, as Detectron2 does not support YOLO. But I'm not surprised that AP is decreased as that is what I also faced in my experiments. I suppose it's possible to get a slight AP improvement by picking the right hyperparameters, though. You could try different training duration or learning rates.
Have you tried newer methods like Unbiased Teacher? Motivated by FixMatch, I believe, it is way better, and they report outperforming CSD nearly 10x on some setups.
My colleague is researching unbiased-teacher
, I will try it soon.
CSD is more simple, but as far as I consider, for semi-supervised, self-training
method maybe more valid.
Just wanted to let you know that I found a bug in my code, and after fixing it the CSD method started to bring improvements of about 1% in many scenarios on VOC. So now I'm more positive about the method's applicability. Sorry for the confusion before.
I implemente csd with scaled-yolov4, but it's AP is decreased. have you plans to try it ?