Closed leedewdew closed 1 year ago
Thank you. I don't think this is related to treating the target boxes at all scales, since it is seen in other detectors such as YOLO. Individual terms of the loss can be adjusted by weighting and tweaking the related hyperparameters (such as obj and obj_pw).
Thank you. I'll try.
Hi, Thanks for your great work. When I transplant this label assign to my code, I found obj loss hard to converge. Will be there many difficult boxes to recognize foreground or background during training due to we treat the target boxes at all scales? In other words, whether this method will introduce many fuzzy prediction to learn?