BCTR: Bidirectional Conditioning Transformer for Scene Graph Generation

2024-07-26

Information Fusion

AI Summary

Scene Graph Generation (SGG) suffers from coarse subject–predicate–object modeling and neglect of bidirectional dependencies between entities and predicates. To address this, we propose a Bidirectional Conditionalized Transformer architecture—the first to enable mutual guidance between visual feature processing and semantic decoding: visual features dynamically modulate semantic decoding, while semantic structures reciprocally constrain visual attention, thereby facilitating joint optimization of objects and relations. Our method comprises a dual-stream conditional encoder, a learnable relational prior module, and a contrastive relational rescoring mechanism. Evaluated on Visual Genome, our approach achieves a +3.2% improvement in Recall@100, demonstrating substantial gains in long-tail relation recognition and modeling of co-occurring multiple relations.

Technology Category

• Computer Vision: Visual Reasoning & Symbolic Representations
• Reasoning under Uncertainty: Relational Probabilistic Models
• Machine Learning: Statistical Relational/Logic Learning

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Problem

• Addresses unidirectional conditioning limitation in SGG
• Proposes bidirectional conditioning for entity-predicate interaction
• Enhances generalization to unseen semantic relationships

Innovation

• Bidirectional conditioning factorization in semantic space
• Multi-stage interactive feature augmentation module
• Random feature alignment with multi-modal knowledge