Part-Aware Segmentation for Fine-Grained Categorization

It is difficult to segment images of fine-grained objects due to the high variation of appearances. Common segmentation methods can hardly separate the part regions of the instance from background with sufficient accuracy. However, these parts are crucial in fine-grained recognition. Observing that fine-grained objects share the same configuration of parts, we present a novel part-aware segmentation method, which can get the foreground segmentation from a bounding box with preservation of semantic parts. We firstly design a hybrid part localization method, which combines parametric and non-parametric models. Then we iteratively update the segmentation outputs and the part proposal, which can get better foreground segmentation results. Experiments demonstrate the superiority of the proposed method, as compared to the state-of-the-art approaches.