DETR VAL mAP 0.0000

MMdetection3.0 Problem: AP values on the DETR validation set are all 0.0000

Conditions:

1. NWPU-VHR-10 data set was selected, including 455 training sets and 195 verification and testing sets.

2. It can be successfully trained on both Faster-rcnn and YOLOv3 models, but during DETR training, the verification index is all 0.0000.

3. Pre-training model is used

Question:

1. All AP values on DETR verification set are 0.0000

2. In addition, it is also a problem when training Co-DETR, that is to say, as long as the DETR based model has this problem.

Is it because the data set is too small? Or do you need to pay special attention to the adjustment of some specific parameters?