Disentangled Feature Network for Fine-Grained Recognition

Abstract

Most of fine-grained recognition researches are implemented based on generic classification models as the backbone. However, it is a sub-optimal choice because the differences between similar categories in this task are so small that the models must capture discriminative fine-grained subtle variances. In this paper, we design a dedicated backbone network for fine-grained recognition. To this end, we propose a novel Disentangled Feature Network (DFN) that gradually disentangles and incorporates coarse- and fine-grained features to explicitly capture multi-grained features. Thus, it promotes the models to learn more representative features that potentially determine the classification results via easily replacing the original inappropriate backbone. Moreover, we further present an optional error correction loss to adaptively penalize misclassification between extremely similar categories and guide to capture fine-grained feature diversity. Extensive experiments fully demonstrate that when adopting our DFN as the backbone, like freebies, the baseline models boost the performance by about 2% with negligible extra parameters on widely used CUB, AirCraft, and Stanford Car dataset.