Help in reproducing pseudo label mIOU

[mustafa1728]

Hi authors, thank you for this interesting work.

I was trying to reproduce the results reported in the paper. After training the classifier, I changed the gen_labels.py file by adding att = cv2.resize(att, (width, height)) (height and width are that of the image) right after line

RCA/gen_labels.py

Line 73 in c69bef3

att = cv2.imread(att_name, 0)

.
this solved issue #5, and I was able to obtain the pseudo-labels.

I tried evaluating the pseudo-labels for the train-list (not train-aug list) with the ground truths using a custom evaluation script. Without EPS I observed mIOU of 0.70 for the pseudo-labels provided by you while only 0.64 for those generated by me. I understand that the discrepancy in the result of your provided pseudo-labels would be resolved by using EPS, but the discrepancy of the pseudo-labels obtained by me that are so low, would not be resolved still.

I request you to please provide additional instructions to reproduce these results that I may have missed or possible mistakes in my approach.

Thank you for your time.

[maeve07]

Hello, thanks for the interests. You can refer to the upsampling code we uploaded, and then generate pseudo labels. But I'm confused about the mIOU without EPS you mentioned.

[911456451]

Hi，I have the same question bro，the miou with OAA is 0.68 but in paper it is 70. Can you solve this problem?