Major bug: satellitepy/evaluate/utils.py/match_gt_and_det_bboxes used obboxes for IoU calculation if obboxes in detections, even if obboxes are missing in ground truth. This leads to an IoU of 0 for all datasets that don't have obboxes (XView, DIOR, Potsdam, VHR) if the model is trained with obboxes, which is always the case for us.
This did not affect the training, but these datasets were completely ignored for testing/evaluation.
Major bug:
satellitepy/evaluate/utils.py/match_gt_and_det_bboxesusedobboxesfor IoU calculation ifobboxesin detections, even ifobboxesare missing in ground truth. This leads to an IoU of 0 for all datasets that don't haveobboxes(XView, DIOR, Potsdam, VHR) if the model is trained withobboxes, which is always the case for us. This did not affect the training, but these datasets were completely ignored for testing/evaluation.