Generating pseudo label is really time-consuming?

[jiangzhengkai]

Hi, I notice generating pseudo labels is really time-consuming? At the same time, the scale for the pseudo label is (1024, 2048) rather than (512, 1024) for the test process.

[BinhuiXie]

Hi, thank you for your comments!
As stated in the paper, self-supervised learning has been used in previous methods such as Tsai et al. 2019, Pan et al. 2020, Wang et al. 2020 and so on, which is usually achieved by iteratively generating a set of pseudo labels based on the most confident predictions on the target data. Our contrastive adaptation strategy is over the feature representations and is orthogonal to the self-supervised learning. Here, we utilize a large resolution because we find it does help increase our performance.