Track new results across pure and applied mathematics.
Search papers, labs, and topics across Lattice.
This paper introduces AutoConcept, a training-free reranker designed for metadata-available composed image retrieval (CIR), which enhances the retrieval process by leveraging concept evidence to create an interpretable memory. By filtering noisy concepts and incorporating metadata-based concept-candidate alignment, AutoConcept achieves significant improvements in early-rank retrieval performance on the FashionIQ dataset compared to existing methods like WeiMoCIR and LinCIR. The findings underscore the value of structured concept memory in enhancing retrieval accuracy beyond traditional query-text matching techniques.
AutoConcept transforms metadata into interpretable memory, significantly boosting image retrieval accuracy without the need for additional training.
Composed image retrieval (CIR) retrieves a target image from a reference image and a text modification. This paper studies metadata-available CIR reranking, where a fixed CIR model first returns a candidate pool and gallery metadata is then used for second-stage concept-guided scoring. We introduce AutoConcept, a training-free reranker that converts concept evidence into an interpretable memory. AutoConcept filters noisy concepts, activates query-relevant positive constraints with an auxiliary negative penalty, and combines base retrieval scores with metadata-based concept-candidate alignment through inference-time calibration. On FashionIQ, AutoConcept yields significant early-rank improvements over WeiMoCIR and consistent plug-in gains on LinCIR candidate pools. Metadata-aware controls show that structured concept memory adds signal beyond direct query-text and extracted-attribute matching, while a query-only variant further supports the effectiveness of concept-level reranking. A supplementary real-human concept-label study indicates that the same memory interface can consume participant-provided evidence. These results position AutoConcept as an interpretable concept-memory reranker for product-style CIR galleries with available metadata.