Abstract
Our work focuses on tackling the challenging but natural visual recognitiontask of long-tailed data distribution (i.e., a few classes occupy most of thedata, while most classes have rarely few samples). In the literature, classre-balancing strategies (e.g., re-weighting and re-sampling) are the prominentand effective methods proposed to alleviate the extreme imbalance for dealingwith long-tailed problems. In this paper, we firstly discover that thesere-balancing methods achieving satisfactory recognition accuracy owes to thatthey could significantly promote the classifier learning of deep networks.However, at the same time, they will unexpectedly damage the representativeability of the learned deep features to some extent. Therefore, we propose aunified Bilateral-Branch Network (BBN) to take care of both representationlearning and classifier learning simultaneously, where each branch does performits own duty separately. In particular, our BBN model is further equipped witha novel cumulative learning strategy, which is designed to first learn theuniversal patterns and then pay attention to the tail data gradually. Extensiveexperiments on four benchmark datasets, including the large-scale iNaturalistones, justify that the proposed BBN can significantly outperformstate-of-the-art methods. Furthermore, validation experiments can demonstrateboth our preliminary discovery and effectiveness of tailored designs in BBN forlong-tailed problems. Our method won the first place in the iNaturalist 2019large scale species classification competition, and our code is open-source andavailable at https://github.com/Megvii-Nanjing/BBN.