Some questions about the CSS work

[Lazzzycoding]

Dear author @WangChangqi98

Thanks for your wonderful work!

I have some questions about this novel work. The first one is idea of utilizing latent space (logit representation) to enhance the S4 task be adapted to the supervised semantic segmentation task. The second question is that the idea of using the logit representation be adapted to the 3D supervised classification or segmentation task.

Thanks a lot.

Best wishes.

[WangChangqi98]

Thanks for following our work. I think using the regularization of latent space as an auxiliary task benefits the supervised semantic segmentation task, and there are some papers discussing about it. (e.g., Exploring cross-image pixel contrast for semantic segmentation)

[WangChangqi98]

In regards to your second question, I'm afraid that I am not familiar with 3D supervised classification or segmentation task, but I guess it is OK to take the contrastive learning as an auxiliary task to train a more powerful segmentation or classification model.

[Lazzzycoding]

Hi @WangChangqi98

Thanks for you kind reply. I also have some questions now. What's the difference between logit space and representation space? How should I define logit space and representation space in a new classification or segmentation space?

Thanks again.

[WangChangqi98]

Hi! In general, logit space is a low dimensional space (the dimension is the same with the segmentation class or classification class), representation space is a high dimensional space (always is 256 or 512 in segmentation)

[Lazzzycoding]

Hi! In general, logit space is a low dimensional space (the dimension is the same with the segmentation class or classification class), representation space is a high dimensional space (always is 256 or 512 in segmentation)

OK. Thanks. I will move forward to understand this paper in depth.

[WangChangqi98]

You are welcome!

[Lazzzycoding]

In your paper, the collaborative space supervision and contrastive learning is conducted between the labeled samples from representation space and unlabeled samples from logit space, right?

[Lazzzycoding]

Hi! In general, logit space is a low dimensional space (the dimension is the same with the segmentation class or classification class), representation space is a high dimensional space (always is 256 or 512 in segmentation)

The representation space could be understood with features from the backbone with a high dimension, and then imported to the logit space to compute loss, right?

[Lazzzycoding]

How to understand the prototype in this paper? What's the difference between prototype and representations?

[WangChangqi98]

How to understand the prototype in this paper? What's the difference between prototype and representations?

Could I ask you questions on Wechat? My wechat is flysnow-hxb. Thanks a lot.

Yes, I sent the application to you.