Data augmentation for label images

[yulei1234]

Thank you for your novel work, I have a question, why not do strong data enhancement on labeled data, I think the quality of such pseudo labels will also be improved. Looking forward for your response, thank you

[kihyuks]

Hi, We have studied the impact of the quality of pseudo labels on the performance of STAC in Section 5.4 and Table 5 of our paper. While the quality of pseudo labels matters, but it was not so significant and the impact was not always positive. Our design of adding strong augmentation only to unlabeled data is in line with that of FixMatch.

We also tried training STAC with strong augmentation on labeled data, but there was no significant performance improvement. This is not included in the paper, but you can easily train STAC with strong augmentation on labeled data by modifying this line:
TRAIN.AUGTYPE_LAB='default' -> TRAIN.AUGTYPE_LAB='strong'