why the learned position attention map highlights regions of the same class as the query pixel

[junfu1115]

why the learned position attention map highlights regions of the same class as the query pixel

[junfu1115]

I think it could enhance intra-class compactness on which softmaxloss has not focused

[yaoqi-zd]

@junfu1115 so you think it is because under pixel-level supervision, pixels of the same class shares similar feature, so the position attention map would highlight regions of the same class as the query pixel ?

[junfu1115]

@yaoqi-zd You could say that