Learning Visual-Audio Dissonance for Moment Retrieval and Highlight Detection

Language-guided video moment retrieval and highlight detection have made impressive advancements recently. Most existing methods primarily focus on visual cues in videos. In reality, audio cues in videos also play an indispensable role. In this paper, we highlight the importance of language-guided visual-audio temporal semantic dissonances and propose a novel framework named VABooster. It aims to leverage audio information as boosting cues to jointly address moment retrieval and highlight detection. We propose a visual-audio synergy network to learn task-specific visual-audio features. Additionally, we introduce an omni-modality interaction component with the dual branch structure to capture visual-audio-text contexts for different tasks. A saliency token adapter is designed to regulate the learning process of global saliency tokens. VABooster surpasses the existing state-of-the-art methods with fewer parameters. Extensive experiments and in-depth ablation studies on QVHighlights, Charades-STA, TVSum, and Youtube Highlights datasets substantiate the effectiveness and robustness of the proposed framework.