Delayed Bidirectional Alignment via Disentangled Audio Semantics for Audio-Visual Segmentation

Audio–Visual Segmentation (AVS) aims to localize sound-producing objects at the pixel level by integrating auditory and visualcues. However, existing methods often struggle with multi-source en-tanglement and audio–visual misalignment, leading to a dominance biastoward acoustically or visually salient objects (i.e., louder or larger ones)at the expense of subtler or co-occurring sources. To address these chal-lenges, we propose DDAVS: Delayed Bidirectional Alignment via Dis-entangled Audio Semantics for Audio-Visual Segmentation. To mitigatemulti-source entanglement, DDAVS employs learnable queries to extractaudio semantics and anchor them within a structured semantic spacederived from an audio prototype memory bank. This process is furtheroptimized through contrastive learning to enhance discriminability androbustness. To alleviate audio–visual misalignment, DDAVS introducesdual cross-attention with delayed modality interaction, improving therobustness of multimodal alignment. Extensive experiments on the AVS-Objects and VPO benchmarks demonstrate that DDAVS achieves state-of-the-art performance across single-source, multi-source, and multi-classmulti-instance scenarios. These results validate the effectiveness and gen-eralization ability of our framework under challenging real-world au-dio–visual segmentation conditions. Project page.