CSMCIR: CoT-Enhanced Symmetric Alignment with Memory Bank for Composed Image Retrieval
Researchers have introduced CSMCIR, a novel framework for Composed Image Retrieval (CIR) that addresses the critical issue of representation space fragmentation in existing methods. Traditional CIR systems struggle with misaligned feature spaces due to heterogeneous modalities processed by distinct encoders. CSMCIR proposes a unified representation approach utilizing three synergistic components: a Multi-level Chain-of-Thought (MCoT) prompting strategy to generate semantically compatible captions, a symmetric dual-tower architecture with shared-parameter Q-Formers for consistent cross-modal encoding, and an entropy-based, temporally dynamic Memory Bank for high-quality negative sampling. This architectural symmetry ensures efficient query-target alignment from initialization. Extensive experiments across four benchmark datasets demonstrate that CSMCIR achieves state-of-the-art performance while offering superior training efficiency. Ablation studies confirm the effectiveness of each component, marking a significant advancement in multimodal retrieval systems by bridging the gap between visual and textual data through enhanced semantic compatibility and structural symmetry.
Wire timeline
CSMCIR: CoT-Enhanced Symmetric Alignment with Memory Bank for Composed Image Retrieval
Researchers have introduced CSMCIR, a novel framework for Composed Image Retrieval (CIR) that addresses the critical issue of representation space fragmentation in existing methods. Traditional CIR systems struggle with misaligned feature spaces due to heterogeneous modalities processed by distinct encoders. CSMCIR proposes a unified representation approach utilizing three synergistic components: a Multi-level Chain-of-Thought (MCoT) prompting strategy to generate semantically compatible captions, a symmetric dual-tower architecture with shared-parameter Q-Formers for consistent cross-modal encoding, and an entropy-based, temporally dynamic Memory Bank for high-quality negative sampling. This architectural symmetry ensures efficient query-target alignment from initialization. Extensive experiments across four benchmark datasets demonstrate that CSMCIR achieves state-of-the-art performance while offering superior training efficiency. Ablation studies confirm the effectiveness of each component, marking a significant advancement in multimodal retrieval systems by bridging the gap between visual and textual data through enhanced semantic compatibility and structural symmetry.
cs.AI updates on arXiv.org