Multi-scale Discriminative Patches for Fined-Grained Visual Categorization.

Fine-grained visual categorization (FGVC) is a challenging vision problem since the similar appearance between object classes. It is important to note that human visual recognition system generally focuses on the specific part to distinguish those confused classes, which is also the breakthrough point for FGVC. In this paper, we will introduce the feedback mechanism of CNN to extract multi-scale discriminative patches. The extracted patches show more significance than the whole object region. Compared with tradition methods, we only require the object-level label rather than part-level annotations. Experiments on Caltech-UCSD Birds-200-2011 demonstrate the effectiveness of our method in solving FGVC.