Dual-referenced assistive network for action quality assessment

Keyi Huang, Yi Tian, Chen Yu, Yaping Huang

Abstract: Highlights•We propose a Rating-guided Attention module, which introduces a set of semantic-level referenced assistants to refine coarse-grained features into rating-informed features. These rating-informed features integrate hierarchical semantic information, which could capture the fine-grained quality-related information that is instructive for score estimation.•We exploit a couple of Consistency Preserving constraints, which apply a group of individual-level referenced assistants to remove redundant features unrelated to the action quality. These constraints guide the model to focus on important features and suppress score-unrelated interference.•The experiments on AQA-7 and MTL-AQA datasets demonstrate the effectiveness of our DuRA network.