GLRM: Graph Learning and Regional Messaging for Unbiased Scene Graph Generation

Scene Graph Generation (SGG) seeks to extract semantic entities and their relationships from images. However, inherent imbalances in existing datasets result in long-tailed predicate distributions and missing entity relationships, leading to an overabundance of background-class annotations. To address these challenges, we proposes the incorporation of pseudo-relationships into entity pairs with missing annotations, thereby enriching the triplet count and mitigating the effects of long-tailed distributions. Building on this strategy, we propose the GLRM network, which comprises a Graph Learning Structure (GLS) module and a Regional Message Passing Network (RMPN). The GLS addresses the problem of missing and irrelevant relationships by learning entity correlations, removing irrelevant edges, and introducing missing edges to allocate pseudo-relationships more accurately. Meanwhile, the RMPN enhances the semantic representation of nodes and edges by leveraging structured representations and multi-round message passing to capture contextual information across regions. Together, these modules enable GLRM to generate pseudo-relationships for background classes, producing unbias scene graphs. Experimental results validate that GLRM significantly enhances the overall performance across various SGG models, effectively addressing the challenges posed by long-tailed distributions and missing relationships.