Efficient Image Embedding for Fine-Grained Visual Classification

Fine-grained visual classification (FGVC) is a task belonging to multiple sub-categories classification. It is a challenging task due to high intraclass variation and inter-class similarity. Most exiting methods pay attention to capture discriminative semantic parts to address those problems. In this paper, we introduce a two-level network which consists of raw-level and object-level networks, and we name it “Efficient Image Embedding”. Its training procedure has two stages which the raw-level is for localization by the aggregation of feature maps, and the last is for classification. The two-level use Adaptive Angular Margin loss (AAM-loss), which improve an intra-class compactness and inter-class variety of image embedding. Our approach is to identify object regions without any hand-crafted bounding-box, and can be trained in an end-to-end manner. It has achieved better accuracy on two datasets compared to the existing work, which are 89.0% for CUB200-2011 and 93.3% for FGVC-Aircraft.