search

dvlab-research / MiSLAS

Improving Calibration for Long-Tailed Recognition (CVPR2021)
MIT License
143 stars 25 forks source link

About the BN part #2

Closed tzhxs closed 3 years ago

tzhxs commented 3 years ago

Hi, thanks for your great work. I am wondering about the BN part, it seems that the methods like "cRT" and "DRW" do update the running mean and variances, right? I can not find the code segment which aims to freeze this part.

zs-zhong commented 3 years ago

Hello! Actually, in the final framework of "cRT", "LWS", "DRS", and "DRW", they do update the running mean and variances in BN. While, in the "tau-norm" and "KNN" decoupling models, they don't update them. In our paper, we focus on the over-confidence issue. Due to the different class number composition ratios for the two stages, we believe the shifted BN could have some effects on the recognition performance and the calibration, which hasn't been discussed in previous two-stage methods. According to our experiment, it's proved.

tzhxs commented 3 years ago

Hello! Actually, in the final framework of "cRT", "LWS", "DRS", and "DRW", they do update the running mean and variances in BN. While, in the "tau-norm" and "KNN" decoupling models, they don't update them. In our paper, we focus on the over-confidence issue. Due to the different class number composition ratios for the two stages, we believe the shifted BN could have some effects on the recognition performance and the calibration, which hasn't been discussed in previous two-stage methods. According to our experiment, it's proved.

So, acturally, the shifted BN is a widely used mehotd for "CRT", "LWS" also fixed the affine part?

tzhxs commented 3 years ago

Is there some more details in the BN part of this paper?

zs-zhong commented 3 years ago

Yes, both "cRT" and "LWS" fix the affine part and update the running means and variances in Stage-2. For more details about the shifted BN, please refer to the experiment part. We also think a novel normalization to effectively deal with long-tailed recognition should be discussed and explored in the future.

tzhxs commented 3 years ago

Yes, both "cRT" and "LWS" fix the affine part and update the running means and variances in Stage-2. For more details about the shifted BN, please refer to the experiment part. We also think a novel normalization to effectively deal with long-tailed recognition should be discussed and explored in the future.

Thanks, look forward to your new work