Query-guided predicate decoupling and prototype approximation learning for scene graph generation

doi:10.1016/j.eswa.2025.129525

Query-guided predicate decoupling and prototype approximation learning for scene graph generation

2025-08-29

0

PRE

AI

DOI：10.1016/j.eswa.2025.129525

Original

Original request for help

Save

Abstract

En 中文

• This work addresses the long-tail distribution problem from both feature decoupling and approximate learning perspectives. • At the feature level, a decoupling method is designed to separate entity and predicate representations. • The homogeneity among same category predicate representations is leveraged to alleviate decision-making pressure at the classification level. • Experiments on multiple datasets demonstrate competitive performance and effectively mitigate the issue of biased predictions.

Query-guided predicate decoupling and prototype approximation learning for scene graph generation

Abstract

Journal

Organization