About Comparisons with CBST

[JoyHuYY1412]

Hi, thank you for your great work. I am currently exploring the semi-supervised semantic segmentation works, and find your work very interesting.

I want to ask a general question, that is, how to understand that the distribution alignment of pseudo-labels is better than class-balancing in CBST?

From my viewpoint, CBST (class balance thresholding method) can be seen as an adaptation of your DAST when assuming the distribution of labeled data is balanced, am I right? So why use the imbalanced labeled data distribution to guide the pseudo-labeling selection is better than a balanced distribution? I find it hard to understand.

Could you please give me some advice? Thank you so much.

(You also proposed Random Sampling in DAST, which is not included in CBST. Maybe that's the point?)

[RuiFeiHe]

Hi, thanks for your attention.

I think it should be emphasized that the task itself is imbalanced, and if we force the distribution of pseudo labels to be balanced for all classes, it would violate the distribution of the task. For example, the proportion of the class road is much greater than the class person, if we force them to be balanced, I think the pseudo-labeled data is much different from the true labeled set of the task.

Hence, for your question "So why use the imbalanced labeled data distribution to guide the pseudo-labeling selection is better than a balanced distribution", I think it is not to use the imbalanced distribution, but to use the true distribution of the task. Our method is more adaptive for the specific task considering the imbalance of the data.