Pic2Word: Mapping Pictures to Words for Zero-shot Composed Image Retrieval
This addresses the high cost of labeling for composed image retrieval, enabling broader applicability in tasks like attribute editing and object composition.
The paper tackles the problem of zero-shot composed image retrieval (ZS-CIR) by proposing Pic2Word, a method that trains without labeled triplets using weakly labeled image-caption pairs and unlabeled images, achieving strong generalization and outperforming supervised methods on benchmarks like CIRR and Fashion-IQ.
In Composed Image Retrieval (CIR), a user combines a query image with text to describe their intended target. Existing methods rely on supervised learning of CIR models using labeled triplets consisting of the query image, text specification, and the target image. Labeling such triplets is expensive and hinders broad applicability of CIR. In this work, we propose to study an important task, Zero-Shot Composed Image Retrieval (ZS-CIR), whose goal is to build a CIR model without requiring labeled triplets for training. To this end, we propose a novel method, called Pic2Word, that requires only weakly labeled image-caption pairs and unlabeled image datasets to train. Unlike existing supervised CIR models, our model trained on weakly labeled or unlabeled datasets shows strong generalization across diverse ZS-CIR tasks, e.g., attribute editing, object composition, and domain conversion. Our approach outperforms several supervised CIR methods on the common CIR benchmark, CIRR and Fashion-IQ. Code will be made publicly available at https://github.com/google-research/composed_image_retrieval.