How to encoder background BPM? about vit-pcm HOT 5 CLOSED

Hi @St1ckyfinger, thank you for your kind comment. I am not sure I understood correctly the first question: K is the cardinality of the categories space which already includes background, i.e. PascalVOC12 has 20 categories and adding 1 for background leads to K=21. By default, we set background as the category with index 0.

Regarding the second question: the only minimization of Loss_mce leads to overfocus on most discriminating regions of objects in each view and classify as background the less discriminating ones, which produces a lot of background false positives. To overcome this problem, we jointly minimize Loss_et, which is meant to minimize the output equivariance under different input transformed views of the same image; preserving equivariance helps in reducing intra-class variance of features, so that background false positives become rarer. We observe that adding Loss_et gives more than 15% miou increment and improve categories encoding and localization, including the background.

I hope this answers to your questions!

from vit-pcm.

Very thanks for your response!
In equation(1), image-level labels of the background category are set to 1 for all images, right?

from vit-pcm.

That is correct. Since we want to localize the background, we assume that there is at least one pixel in every image of background/unknown, so we set the background to 1 for all images.

from vit-pcm.

Got it, thanks!

from vit-pcm.

You are welcome :)

from vit-pcm.