Data-efficient generalization for zero-shot composed image retrieval

doi:10.1016/j.patcog.2026.113187

Data-efficient generalization for zero-shot composed image retrieval

2026-01-28

0

PRE

AI

DOI：10.1016/j.patcog.2026.113187

Original

Original request for help

Save

Abstract

En 中文

• Observe modality discrepancy and distribution shift issue in ZS-CIR task • Enhance pseudo-word token with linguistic semantics • Excavate zero-shot capability of VLMs with novel loss functions • Achieve superior performance on four benchmarks with less training data • Achieve an optimal balance between training and inference cost

Data-efficient generalization for zero-shot composed image retrieval

Abstract

Journal

Organization