New dynamic graph neural network with adaptive feature selection improves RGB-D indoor scene recognition by intelligently selecting key features from color and depth data.

Proposes a dynamic graph model that adaptively selects key local features from both RGB and depth modalities for better scene understanding

Uses a graph-based approach to model relationships between objects and scenes in 3D indoor environments

Organizes selected nodes into three hierarchical levels to represent near and far spatial relations among objects

Addresses the open problem of effective exploitation of multi-modal local features in indoor scene recognition tasks