Synergistic Space-Vision Processing for Predicate Inference

Scene graph generation (SGG) aims to parse an image into a structured graph of objects and their predicates, enabling explicit relational reasoning for visual understanding. However, prevailing methods often over-predict geometric predicates, resulting in scene graphs that are factually correct yet semantically shallow. While recent works effectively attribute this phenomenon to the long-tailed data distribution, we identify another critical factor driving such biased prediction: co-occurrence-induced representation entanglement, where geometric and non-geometric predicates that frequently co-occur are encoded into overly similar representations. To this end, we introduce Dual-stream Synergistic Network (DS-Net) that models geometric and non-geometric predicates with two specialized streams, coupled with a bidirectional cross-stream fusion mechanism. The space stream focuses on spatial and structural cues, while the vision stream captures fine-grained visual evidence and semantic priors. Extensive experiments show that DS-Net consistently improves predicate inference, achieving 1.3\% $\sim$ 6.1\% absolute gains in mR@100 on the SGGen task when integrated into existing SGG methods. These results highlight the importance of synergistic modeling of geometric and non-geometric predicates for generating semantically richer scene graphs.

Artificial intelligence systems utilize scene graph generation to parse images into structured networks of objects and their predicates, enabling explicit relational reasoning. However, prevailing methods frequently overpredict basic geometric relationships, producing outputs that are factually correct but semantically shallow. We discovered that a critical factor driving this biased prediction is information conflation, where geometric and non-geometric predicates that frequently appear together are encoded into overly similar representations. To address this issue, we introduce the Dual Stream Synergistic Network, which models these predicates through two specialized pathways coupled with a bidirectional fusion mechanism. The space stream focuses entirely on spatial and structural cues, while the vision stream captures fine-grained visual evidence and semantic priors. Extensive experiments demonstrate that our network consistently improves predicate inference when integrated into existing methods, achieving significant absolute gains in core evaluation metrics. These findings highlight the importance of synergistic modeling to accurately identify tail predicates and generate semantically richer scene graphs for comprehensive visual understanding.