Adaptive selection and exploitation of key local features for RGB-D indoor scene recognition

Develop adaptive methods that select and effectively exploit key local features from both RGB images and depth maps for RGB-D indoor scene recognition, with the goal of improving classification performance over approaches that rely on global features or fixed local feature sets.

Background

RGB-D indoor scene recognition leverages complementary information from RGB images and depth maps. While global features capture overall scene context, numerous studies have shown that local features are critical for handling cluttered objects and complex spatial layouts typical of indoor scenes.

The paper emphasizes that, despite progress, determining how to adaptively select and utilize the most informative local features from both modalities remains unresolved. This stated open problem motivates their proposed dynamic graph neural network with an adaptive node selection mechanism designed to identify and exploit such features effectively.