Dimensions overall score 9.0

• MCoT-MVS integrates attention-aware multi-level vision features guided by reasoning cues from a multi-modal large language model (MLLM) for Composed Image Retrieval.method 95%
• The method leverages an MLLM to perform chain-of-thought reasoning on multimodal input, generating retained, removed, and target-inferred texts.method 95%
• Textual cues from MLLM reasoning guide two reference visual attention selection modules to extract discriminative patch-level and instance-level semantics from the reference image.method 95%
• A weighted hierarchical combination module is designed to fuse multi-granular visual cues with modified text and imagined target description.method 95%
• Extensive experiments on CIRR and FashionIQ benchmarks demonstrate that MCoT-MVS consistently outperforms existing methods and achieves new state-of-the-art performance.result 95%
• Existing methods often struggle to extract correct semantic cues from reference images that reflect user intent under textual modification prompts, resulting in interference from irrelevant visual noise.limitation 90%
• The approach aligns composed queries with target images in a unified embedding space.technical 90%
• Code and trained models for MCoT-MVS are publicly released.technical 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card