Mask-RCNN merge two predicted masks of the same class with overlap

Question

i`m working on a prediction modell for crack segmentation based on matterports Mask-RCNN implementation and to this point my modell managed to mask cracks relatively precise