Object-Aware Attention Branch Network for Interior Style Scene Recognition
Kentaro Fukao, Keiko Ono, Yuki Tani · 2023
Recently, there has been a growing interest in developing recommendation systems that capture users' preferences and interests, while focusing on visual styles has gained increasing attention. However, understanding interior-style scenes is challenging because of the complex interplay of various objects that must be correctly interpreted. To properly understand interior-style scenes, extracting appropriate image features and providing visual explanations for the objects that determine the style is necessary. This study proposed a model that can simultaneously extract image features at multiple scales and provide visual explanations for the objects that characterize each class. Specifically, we adopted a hierarchical attention branch network (ABN) for visual explanations and applied atrous convolution to the feature extractor. We hypothesized that atrous convolution could extract various specific objects of each class that differ in size because its algorithm extracts features at an arbitrary resolution. Our evaluation results show that a conventional hierarchical ABN could not extract interior objects accurately, and our proposed model can detect specific objects by incorporating atrous convolution.