search

WangChangqi98 / CSS

[ICCV'23] Space Engage: Collaborative Space Supervision for Contrastive-based Semi-Supervised Semantic Segmentation
33 stars 3 forks source link

Question about the loss #3

Closed Hugo-cell111 closed 1 year ago

Hugo-cell111 commented 1 year ago

Hi! I wonder why the supervised loss function is CE Loss when the dataset is Pascal VOC while the function is ProbOhemCrossEntropy2d when the dataset is the other one (such as Cityscapes and COCO)? Thanks!

WangChangqi98 commented 1 year ago

Hi, this is because the ProbOhemCrossEntropy2d can achieve better performance especially on Cityscapes. Meanwhile, the SOTAs which we compared with are also adopted ProbOhemCrossEntropy2d on Cityscapes, so for a fair comparison, we need to use it.

Hugo-cell111 commented 1 year ago

Yes, I know the ProbOhemCrossEntropy2d is always used in the previous works. Do you mean even using the normal CE Loss will achieve the best performance, so the ProbOhemCrossEntropy2d is not needed in VOC?

WangChangqi98 commented 1 year ago

Yes, I think so, ProbOhemCrossEntropy2d may be used to find more hard pixels but in VOC, CE may be enough to reach the convinced performance.

WangChangqi98 commented 1 year ago

Unfortunately we did not try to apply ProbOhemCrossEntropy2d on VOC, and this may be a good try to see whether ProbOhemCrossEntropy2d can achieve better performance or not.

Hugo-cell111 commented 1 year ago

Thanks!