Open nemonameless opened 5 years ago
Hi, we have not changed the data representations in maskrcnn benchmark. We only add a MaskIoU head and we have not changed others' modules. Our experiments show that there is little effect on loss of box head and mask head, so it may not degrade the box AP so much. You need to check whether the maskiou head you added affects the training of the box head.
Have you reproduce this on mmdetection? @zjhuang22 We mainly follow your code and modify mmdetection but get 1.5 reduction in detection and no gain on ins seg compared with mask rcnn on mmdetection. What are the differences between the data representations of the two frameworks,is there any points for attention ?