Fine-Grained Recognition with Incremental Classes

This work focuses on dealing with fine-grained recognition problems when incremental classes emerge. The task is desirable to not only distinguish subordinate visual classes based on discriminative but subtle object parts, but also recognize new coming sub-classes without suffering from catastrophic forgetting. In this paper, we first propose to localize both object- and part-level image regions for capturing powerful fine-grained patterns. Then, these fine-grained regions are fed into a bilateral network consisting of a stable branch and a flexible branch for supporting observed and incremental sub-classes recognition respectively. Moreover, a cumulative adaptation strategy is further equipped to adjust the network training during the incremental sessions. Meanwhile, to better retain the modeling capability of observed classes, we also replay samples from previous classes by a hallucination approach. Experiments are conducted on three popular fine-grained recognition datasets and results of the proposed method can reveal its superiority over state-of-the-arts.