How can I improve the mask prediction by Mask RCNN model? Bounding box and class prediction seems to be okay in my case. Masks are kind of acceptable for small objects but not f
