COMBINER: Composed Image Retrieval Guided by Attribute-based Neighbor Relations
Title: COMBINER: Leveraging Attribute-Based Neighbor Relations for Composed Image Retrieval
Abstract
Composed Image Retrieval (CIR) poses a significant challenge in the field of information retrieval, aiming to locate specific images using combined multimodal inputs. While CIR methodologies have advanced recently, existing approaches frequently fail to adequately handle scenarios where images share visual similarities but possess distinct attributes. This oversight can negatively impact both the fusion of multimodal features and the accuracy of similarity modeling. To address this gap, we propose a unified representation for cross-modal features that relies on attribute prototypes.
However, resolving these issues is complex due to three primary challenges: semantic entanglement at the attribute level, modality inconsistencies, and a lack of supervised signals. To overcome these hurdles, we present COMBINER, a network for COMposed image retrieval guided By attrIbute-based NEighbor Relations. Our framework incorporates three key components:
- An Adaptive Semantic Disentanglement module that separates attribute features derived from multimodal primitive features.
- A Unified Prototype-based Composition module designed to construct Cross-modal Unified Prototypes (CUP) and streamline the composition of multimodal features.
- A Dual Relations Modeling module that extracts both pairwise and neighbor relationships based on attribute similarity.
COMBINER distinguishes itself as the first study to specifically address the phenomenon of samples that are visually similar yet attribute-unrelated, a limitation in traditional neighbor-relation-based CIR methods. By utilizing an attribute prototype-based similarity metric, our approach enables a more precise interpretation of semantic relationships among samples. Extensive experiments across three benchmark datasets validate the effectiveness of COMBINER. The source code for our method will be made available at https://github.com/Lee-zixu/COMBINER.
Source: arXiv Generated at: 2026-06-04 00:00:00 UTC
