Scene Recognition With Comprehensive Regions Graph Modeling

Learning the regional contents of scenes comprehensively is key to scene recognition. Due to semantic diversity and spatial complexity in scene images, modeling based on these regional contents is challenging. The current works mainly focus on some small and partial regions of the scene, while ignoring the majority region of the scene. In contrast, we propose the Semantic Regional Graph modeling framework for the comprehensive selection of discriminative semantic regions in scenes. To explore the relations of these regions, we propose to model these regions in geometric aspect based on the graph model, and generate the discriminative representations for scene recognition. Experimental results demonstrate the effectiveness of our method, which achieves state-of-the-art performances on MIT67 and SUN397 datasets.