Configurable Graph Reasoning for Visual Relationship Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Knowledge Representation

Background:

• Visual commonsense reasoning faces challenges with long-tailed visual relationships and biases in object/relation labels.
• Current methods use fixed reasoning paths ({subject, object → predicate}) for infrequent relationships, but this is limited by dataset bias and semantic gaps.

Purpose of the Study:

• To propose Configurable Graph Reasoning (CGR) for decomposing visual relationship reasoning paths.
• To enable configurable knowledge selection and personalized graph reasoning for each relation type and image.

Main Methods:

• CGR decomposes the reasoning path and external knowledge incorporation.
• It learns to match and retrieve knowledge for subpaths from a commonsense knowledge graph.
• Knowledge is selectively composed to form a knowledge-routed path.

Main Results:

• CGR adaptively configures reasoning paths, bridging the semantic gap between commonsense knowledge and real scenes.
• Achieves improved knowledge generalization and consistently outperforms state-of-the-art methods on benchmarks.
• Demonstrates explainable and compelling configurations of reasoning paths.

Conclusions:

• CGR offers a flexible and effective approach to visual commonsense reasoning.
• The method shows strong performance and adaptability with different knowledge graphs.
• It advances the field by enabling more robust and generalizable visual relationship recognition.