Spatial-Information Enhanced Graph Transformer for Scene Graph Generation

Scene graph generation (SGG) is a fundamental task in visual understanding that aims to detect objects and predict their pairwise relationships in a structured manner. While recent approaches based on graph neural networks (GNNs) have shown promising results, they remain limited in their ability to effectively exploit fine-grained geometric relationships between objects and capture long-range contextual dependencies. To address these limitations, we propose a Spatial-Information Enhanced Graph Transformer for SGG. Our approach builds upon the Graphomer framework to model relational context and incorporates a novel spatial fusion module that explicitly encodes geometric relationships between object bounding boxes—such as Euclidean distance, relative direction, and area overlap. These spatial features are integrated into the encoder, providing enriched positional priors for downstream Transformer layers. The proposed model enables more accurate and globally consistent relation predictions. Extensive experiments conducted on the Visual Genome benchmark demonstrate that our method achieves superior performance over strong baselines, particularly in scenarios with strong spatial dependencies and complex multi-object interactions, highlighting the effectiveness of the proposed spatial encoding strategy in enhancing visual understanding.