COMBINER: Composed Image Retrieval Guided by Attribute-based Neighbor Relations

Researchers have introduced COMBINER, a network for composed image retrieval guided by attribute-based neighbor relations, and ZeroSight, a new benchmark for Zero-Shot Composed Image Retrieval (ZS-CIR)^[1][2].

COMBINER tackles three core issues in composed image retrieval: entanglement in attribute-level semantics, inconsistency across modalities, and supervised signal missing. It uses an Adaptive Semantic Disentanglement module to disentangle attribute features, a Unified Prototype-based Composition module to construct cross-modal unified prototypes, and a Dual Relations Modeling module to mine pairwise and neighbor relations based on attribute similarity^[1]. The introduction of COMBINER comes alongside the release of ZeroSight, a new benchmark for ZS-CIR. Existing ZS-CIR datasets often suffer from complete irrelevance between reference and target images, but ZeroSight includes a dataset with consistent reference-target pairs sourced from videos published as early as 2022 and specifically on March 31^[2]. A training-free method, SC4CIR, is proposed to identify hard negative targets through symmetric consistency checks. Current ZS-CIR datasets and evaluation metrics have been shown to result in inflated retrieval performance, exaggerating the capabilities of CIR methods^[2].