Class Activation Values: Lucid and Faithful Visual Interpretations for CLIP-based Text-Image Retrievals

Transformer-based text-image matching model, known as CLIP, has garnered significant attention owing to its exceptional performance in text-image retrieval tasks and downstream applications. However, the interpret-ability of CLIP remains underexplored. Existing interpretation methods for Transformers often struggle with incomplete and unreliable attributions within the image and text modalities, respectively. In this paper, we propose a fine-grained interpretation method, termed Class Activation Values (CAV), to provide lucid and faithful visual explanations for CLIP-based text-image retrievals. Specifically, we systematically perform multi-scale accumulation and fusion of class-specific gradients and activation value features to generate high-definition explanations for the image encoder. Furthermore, we present element-wise gradient-based weights to attribute fine-grained relevance between value features and output similarity within the text encoder. The proposed CAV is capable of simultaneously rendering detailed and credible explanations due to its precise feature attribution. Extensive qualitative and quantitative experiments are conducted on the ImageNet-1k and MS COCO datasets, and the experimental results demonstrate that CAV outperforms state-of-the-art interpretation methods in both faithfulness and localization assessments across image and text modalities.