Purpose allow_negative_crop in train_pipeline of MM Grounding DINO

[AwePhD]

Hello,

In the base config of MM Grounding DINO, the train_pipeline has a RandomCrop with its allow_negative_crop set to True. Namely, the pipeline can produce an image with no bounding boxes on it.

My question is: why is it there? The BaseDataset.__getitem__ already use a refetch in case the transform return None. I do not see the point to use compute time on frame with no annotations.
Anyone can help me to understand this settings please?

Have a nice day,
Mathias