Question about pseudo label and category mismatch. about sigma HOT 2 CLOSED

Hi, thanks for your good questions. I am happy that you have such deep and insightful thoughts and show interest in our work.

Pseudo Label
The smoothness property of the graph operations will make it robust for noisy labels. Therefore, we set a very low threshold (0.5) to sample nodes. More insights can be found in several noisy-label-related works, e.g., the famous graph propagation in pseudo-labeling (semi-supervised learning) here.

To further improve the performance, you can consider other scenarios related to noisy labels to find the solution. E.g., introduce Teacher-Student EMA, add CycleGAN interpolated data (Note that this will significantly improve the self-training-based method), and logit engineering.

Category Mismatch
The primary function of DNC is to enable the unbiased adaptation signal in each iteration, designed for the cross-domain inconsistent class distribution. Yes, we also found that it will make a more noticeable effect in the scenarios with more classes, e.g., P2C. Moreover, this can be seen as a fine-grained version of the existing prototype-based approach [1,2].

[1] Cross-domain Object Detection through Coarse-to-Fine Feature Adaptation
[2] RPN Prototype Alignment for Domain Adaptive Object Detector

Hope this will help.

from sigma.

Thanks for your timely reply. Your response is really helpful to me!

from sigma.