It is possible to replace the backbone of a Mask R-CNN by a self-supervised CNN?

Question

I would like to implement self-supervised learning (more specifically: image colorization) to see if it can improve instance segmentation of vegetation in top-down (drone) i