peiwang062
commented
2 years ago The segmentation mask is not a must. On COCO, we basically followed the setting of point DETR paper, '(a) if the object has instance segmentation, randomly sample a point from the instance mask as the point annotation for the object; (b) if not, simply randomly sample a point in its bounding box.' On other datasets without mask annotation, we simply sampled a point as (b). So please refer to the split file of other datasets or modify coco splitting code as needed.
Dear Authors,
I have a specific question about split_dataset_coco_omni.py
I do not have any segmentations in my annoatation. Only bbox. The code does not work if I don't have segmentation. Is other scripts and the code written under the assumption that segmentations exist? I can convert bbox to segmentation but that should not have been needed as the code should support bbox only.
My ann['segmentation'] is empty: that is [] But then I get out of range error during annToMask.
Please help.