search

Chen-Ziyang / VPTTA

Code for [CVPR 2024] Each Test Image Deserves A Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation.
40 stars 6 forks source link

The Question About the Usage of Source Domain Data #4

Closed lixiang007666 closed 7 months ago

lixiang007666 commented 7 months ago

According to my understanding, in the Statistics Alignment step, this method still utilizes the source domain data and its labels.

So, it's not entirely source-free, is it?

Chen-Ziyang commented 7 months ago

Thanks for this concern. In the inference phase, we introduce the source statistics by using the running_mean and running_var stored within the BN layers. Therefore, there is no need to utilize source domain data or its labels during inference. The related code can be found in utils/convert.py.