Annotations Mismatch in Datumaro dataset

[kprokofi]

Annotations mismatch in Datumaro dataset. Instance Segmentation training fails due to number of boxes and masks mismatch. Moreover # annotations in dm_subset != # annotations in json file

Steps to Reproduce

1. Take real dataset to train on (wgisd in my case, I will provide env and a dataset)
2. Run normal training: otx train --config src/otx/recipe/instance_segmentation/maskrcnn_r50_tv.yaml --data_root ../../detection/wgisd_merged_coco_small
3. Investigate dataset in OTXInstanceSegmentation Dataset
4. Ensure that all annotations are correctly loaded, compare with original json file

Environment:

• OS:
• Framework version:
• Python version:
• OpenVINO version:
• CUDA/cuDNN version:
• GPU model and memory:

[kprokofi]

So, the current behavior is similar to the old one right now?
Are all pieces of polygons merged for training?

[kprokofi]

I am validating now performance with updated Datumaro on the develop branch

[kprokofi]

I can confirm, that:

1. there is no anymore mismatch in number of objects for Masks and Boxes
2. Accuracy for MaskRCNN-TV looks fine
3. mAP for RF-DETR matches mAP achieved on "old" develop

Hanging / freezing before training starts still preserved

[AlbertvanHouten]

Thanks for confirming. Hanging/freezing will be further investigated in #5352