Question about TripleMixer's mIoU performance on SemanticSTF in Table VIII

[zs3d]

Hello, I noticed in Table VIII that the mIoU performance of TripleMixer on the SemanticSTF dataset is relatively low. I would like to ask whether this is due to the denoising model being trained on the Weather-KITTI dataset, which leads to suboptimal cross-domain denoising performance? Or is it simply because the segmentation model was trained directly on the SemanticKITTI dataset, and its performance on SemanticSTF is relatively poor even without any noise interference? Looking forward to your response. Thank you very much!