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 initialization of prototypes #4

Closed Hugo-cell111 closed 1 year ago

Hugo-cell111 commented 1 year ago

Hi! I find that the prototypes are initialized as the torch.zeros Tensors, and I wonder whether the noise will be brought by the representation space pseudo labels? Thanks!

WangChangqi98 commented 1 year ago

Hi, the first prototype in the training process is the prototype calculated in the first mini-batch, you can find it in ./generalframeworks/loss/loss.py, and there will be some noise in pseudo-labels from the representation space, we follows prior works and use a simple threshold sampling strategy to improve the quality of the pseudo-labels and meanwhile, many other approaches can be used to polish the pseudo-labels from the representation space.

Hugo-cell111 commented 1 year ago

I see. Thank you!