I’ve coded YOLOX from sctrach, but I have a problem with the objectness loss (IoU branch). In the paper they not report how the loss is computed. However, searching online I found this formula
YOLOX total loss
The problem is that, if for regression and class loss just positives prediction are used, for the objectness, all the predictions, including the negatives, should be used. So dividing by just the number of positives should not be correct in this case, and will lead to huge objectness loss values ?