Towards Alleviating Text-to-Image Retrieval Hallucination for CLIP in Zero-shot Learning

TL;DR

This paper identifies and addresses the hallucination problem in CLIP for zero-shot text-to-image retrieval, proposing a novel method called BSAP that improves retrieval accuracy by using auxiliary prompts and normalization.

Contribution

The paper introduces BSAP, a new approach with auxiliary prompts and normalization to reduce hallucinations in CLIP for zero-shot retrieval tasks, enhancing performance.

Findings

BSAP increases CLIP's performance by 20.6% on RefCOCO.

The method is effective on REC and RIS tasks.

Applicable to other models like ALBEF and BLIP.

Abstract

Pretrained cross-modal models, for instance, the most representative CLIP, have recently led to a boom in using pre-trained models for cross-modal zero-shot tasks, considering the generalization properties. However, we analytically discover that CLIP suffers from the text-to-image retrieval hallucination, adversely limiting its capabilities under zero-shot learning: CLIP would select the image with the highest score when asked to figure out which image perfectly matches one given query text among several candidate images even though CLIP knows contents in the image. Accordingly, we propose a Balanced Score with Auxiliary Prompts (BSAP) to mitigate the CLIP's text-to-image retrieval hallucination under zero-shot learning. Specifically, we first design auxiliary prompts to provide multiple reference outcomes for every single image retrieval, then the outcomes derived from each retrieved image in conjunction with the target text are normalized to obtain the final similarity, which alleviates hallucinations in the model. Additionally, we can merge CLIP's original results and BSAP to obtain a more robust hybrid outcome (BSAP-H). Extensive experiments on two typical zero-shot learning tasks, i.e., Referring Expression Comprehension (REC) and Referring Image Segmentation (RIS), are conducted to demonstrate the effectiveness of our BSAP. Specifically, when evaluated on the validation dataset of RefCOCO in REC, BSAP increases CLIP's performance by 20.6%. Further, we validate that our strategy could be applied in other types of pretrained cross-modal models, such as ALBEF and BLIP.