BBN: Bilateral-Branch Network with Cumulative Learning for Long-Tailed Visual Recognition
Boyan Zhou
Quan Cui
Xiu-Shen Wei
Zhao-Min Chen
Megvii Technology
Waseda University
Nanjing University
2020
IEEE/CVF Conference on Computer Vision and Pattern Recognition
1-14
This paper introduces an approach of learning the representation and classifer together based on the same weights but use two different branches to supply examples. The conventional learning branch is the typical long-tail distribution with uniform sampling. The re-balancing branch is a reversed sampler. These are combined together in the cumulative learning block where some hyper parameter alpha is varied such that in the beginning the examples from the conventional branch are used and as epochs go by they are used less and less as the re-balancing branch examples are used. Their cumulative learning strategy is outlined in more detail but alpha depends on the epoch number and the maximum epoch number to be calculated. Their results show great success.