low mIOU in train with coco and ResNet as backbone

[zwtburger]

hi, thank you for your code, I trained the network use coco dataset and ResNet as backbone with 40 epoches, the problem is that Val mIOUs are same value and low (about 0.03x). I just run train_coco.py, are there any wrong on my work? Thanks.

[guanghuixu]

I met the same problem， do you have any idea?

[dfan]

The issue is with the way images are padded in the RandomScaleCrop data augmentation, when the shorter side after rescaling is not large enough to be cropped. As is, the code pads with 0s, but 0 is one of the MS COCO classes evaluated on. This artificially inflates the number of examples of class-0 pixels. You should pad with say 255, and ignore 255 when calculating the cross entropy loss. Also, I believe the starting learning rate of 0.01 is too high - try 0.001. Hope this helps! (I get mIoU of ~0.4673 on the COCO validation set after just 3,000 iterations)

[Qirui-Y]

@dfan Thanks for your plan, I also encountered the same problem, can you provide more detailed ?