Some Instance Segmentation COCOEvaluator mAP scores are suspiciously round numbers (e.g. 90.000 mAP) or "NaN"

[Robotatron]

Everything works fine with the public COCO dataset. However when using my custom dataset for instance segmentation in a polygon format, when training with Mask2Former some class mAP scores are suspiciously round numbers (e.g. 90.000 mAP or 50.0000) or "NaN".

Default trainer used with the COCOEvaluator:

1. Plotting predictions of the model works fine - every class is plotted correctly and shows the model has learned well
2. Plotting the dataset shows all labels of all classes being correctly plotted by Detectron2 -> no visible issue with the annotations
3. I used Shapely Python package to check the polygons for geometric correctness (polygon.is_valid()) -> every annotation is a valid polygon.

I would be glad to hear how can I pinpoint the issue with the COCOEvaluator