BBN: Bilateral-Branch Network with Cumulative Learning for Long-Tailed Visual Recognition

• DOI: 10.48550/arxiv.1912.02413
• Publication Date: 2019-01-01
• Authors: Boyan Zhou; Quan Cui; Xiu-Shen Wei; Zhao-Min Chen
• Keywords: Feature Learning; Benchmark (surveying)

Our work focuses on tackling the challenging but natural visual recognition task of long-tailed data distribution (i.e., a few classes occupy most data, while have rarely samples). In literature, class re-balancing strategies (e.g., re-weighting and re-sampling) are prominent effective methods proposed to alleviate extreme imbalance for dealing with problems. this paper, we firstly discover that these achieving satisfactory accuracy owe they could significantly promote classifier learning deep networks. However, at same time, will unexpectedly damage representative ability learned features some extent. Therefore, propose unified Bilateral-Branch Network (BBN) take care both representation simultaneously, where each branch does perform its own duty separately. particular, our BBN model is further equipped novel cumulative strategy, which designed first learn universal patterns then pay attention tail gradually. Extensive experiments four benchmark datasets, including large-scale iNaturalist ones, justify can outperform state-of-the-art methods. Furthermore, validation demonstrate preliminary discovery effectiveness tailored designs in method won place 2019 large scale species classification competition, code open-source available https://github.com/Megvii-Nanjing/BBN.