search

jiequancui / ResLT

ResLT: Residual Learning for Long-tailed Recognition (TPAMI 2022)
https://arxiv.org/pdf/2101.10633.pdf
MIT License
57 stars 5 forks source link

About the classifiers #1

Closed mingliangzhang2018 closed 3 years ago

mingliangzhang2018 commented 3 years ago

Thanks for proposing such interesting method. I am wondering three branches have diffierent classifiers or same? Furthermore, do you compare with mulit-experts methods as [1]? [1] LONG-TAILED RECOGNITION BY ROUTING DIVERSE DISTRIBUTION-AWARE EXPERTS, 2021 ICLR.

jiequancui commented 3 years ago

Hi,

We use only the shared same classifier. Our method is complementary to the ensemble-based method [1]. A careful comparison will be added in our new version.

mingliangzhang2018 commented 3 years ago

Could you describe in detail how to divide different classes according to different branches? Each group has the same imbalance factor? If one bach size has no tail classes, the tail branch would do not work? Thank you very much.

jiequancui commented 3 years ago

Hi,

Please refer to Sec. 3 in the paper for assigning classes to different branches. If one batch has no tail classes image in the training process, the L{t} will be 0, but L{all} will work normally.