IS the Attention maps for all VOC images released is the final version?

[zmbhou]

I am trying to reproduce to proxy GTs according to the Algorithm 1 by using the DSS and the released attention maps, I find the quality of the generated proxy GTs is poor. and the segmentation performance is far way from that reported. Maybe it is because the attention maps were saved via PNG format and some thing goes wrong? Would you release the code for generating the proxy GTs or the original attention files?

[houqb]

You should ignore those regions in which attentioin values for some pair of classes are both large. For example, for location (i, j), A_{bike} = 0.7 and A_{person} = 0.8. In this case, we do not compute the segmentation loss for it.

Our another paper is under review at this time. We will release the code once it is accepted.

[zmbhou]

Thanks for your reply. Does there exist ignored regions for images with single class? The proxy GTs with single class are of satisfactory as well?

[houqb]

You can set regions with low saliency values (say < 32) to ignored regions. After all, the goal is to make the precision as high as possible. So, sometimes ignoring some regions probably with wrong labels is helpful for gaining better mIoU.

[zmbhou]

Thanks for your reply. I am now trying to reproduce your results as as a baseline for comparison in my revised paper, required by the reviewers. However, the proxy GTs generated by myself according to algorithm 1 can not reproduce the segmentation performance at all. Could you share the proxy GTs with me?

[houqb]

You email address? Or I can provide you with directly the specific numbers.

[zmbhou]

Thanks for your generous help!! My email address is [email protected]. It is better if you can provide the original proxy GTs, because my method focuses on refining the proxy GTs.

[hdjsjyl]

@Andrew-Qibin Can I have the original proxy GTs as well? [email protected], thanks

[houqb]

Done!

[XuYunqiu]

Hi @Andrew-Qibin , would you mind to share your CAM results on VOC 2012 val and test set. I really want to follow your work based on the CAM generated by your proposed method.