AP performance difference

[Iamal1]

• Still confused about #36. I check the msrcnn-res101.log and find the AP there are

OrderedDict([('bbox', OrderedDict([('AP', 0.4012106138083879),'segm', OrderedDict([('AP', 0.3740597044190809) which are different in the paper(Table1 and Table2). So do you mean that the results in paper come with additional tricks ?

• My model uses 2x schedule, mdcn(dcn_v2), panet(ResNeXt101-fpn), multi-scale training, gets 0.480 box AP and 0.426 mask AP, after the same training procedure with maskiou head, finally I get 0.476 box AP and 0.421 mask AP. As the paper suggests, there is an expectation of about 1 point improvement on mask AP without hurting the box AP, I check the losses, the loss for RPN is lower, but mask loss is almost the same. I train and test on coco2017, #46 says the variance is only 0.1-0.2, not that in my case. Whatever the tricks are, the model should be expected to get improvement by introducing maskiou head, as long as the settings are the same. My question is, am I missing some key settings so that the performance is not as good as expected, please inform me if I do.

[zjhuang22]

I think you can reproduce our results using the maskrcnn baseline (without any tricks) to check whether there is the problem of the codes.

[Iamal1]

Many thanks! I'll check it if I had time