Mask Refine in MMDetection

[TonyCongqianWang]

MMYolo lists MaskRefinement as a trick to improve (BBox) detection accuracy. In the reported scores, using masks to refine the training process significantly improves the box accuracy. I am wondering, why there is no such improvement when looking at the scoes for RTMDet. The box AP scores for the ins models are slightly lower.

One reason could be, that the model also outputs the masks and needs more parameters to achive the same accuracy. In that case, it would be nice to have the option to train only for the boxes, but use the masks as refinement - just like in mmyolo? Is that possible or is it restricted due to licensing?