Option to use background-only images during training with MASK-RT-DETR

[Wavelet303]

问题描述 Please describe your issue

I am using MASK-RT-DETR for instance segmentation with a COCO-format dataset.
In my case, the dataset contains a large percentage of background-only images (i.e., images without annotations).

Currently, I noticed that:

During dataset checking/validation, these images are ignored.

During training, they are also not used.

However, for my use case, it is important to include these background-only images in training, since they can help reduce false positives.

👉 Question:
Is there any configuration option (or code modification) that allows MASK-RT-DETR to use background images without annotations during training?
If not, could such an option be added?

This would be very useful in scenarios where negative samples (background) are important to improve model robustness.