Employing ConsistentTeacher to rotated object detection #68
Description
Hi, thank you for your excellent work! I have been trying to employ ConsistentTeacher to rotated object detection. I made some modifications to your code and I trained it successfully on DOTAv1 dataset. However, the sup_loss_bbox, unsup_loss_bbox, unsup_num_gts and unsup_gmm_thr are always 0 during training. What's more, although unsup_loss_cls are not 0, it was really small all the time. And after I trained for 180000 iterations, the detection results are 0 for all the categories. Are there any chance that you can help me to reveal the reason for the above zeros in training and finally get a correct rotated version of ConsistentTeacher? Thank you very much!