DecomCAM: Advancing Beyond Saliency Maps through Decomposition and Integration

TL;DR

DecomCAM addresses the challenge of interpreting complex vision-language models by decomposing class-discriminative activation maps with gradient-guided channel selection and Singular Value Decomposition to produce Orthogonal Sub-Saliency Maps (OSSMs). These OSSMs are integrated using forward-score differences on Gaussian-blurred references, yielding robust, fine-grained saliency maps that emphasize distinct visual attributes while suppressing noise. Extensive experiments on six CLIP-based benchmarks demonstrate superior localization, causal interpretability, and attribute-aligned OSSMs, outperforming prior CAM and decomposition methods and enabling category-level interpretation. The approach offers a scalable, efficient pathway toward transparent, fine-grained understanding of deep models in open-world vision-language tasks.

Abstract

Interpreting complex deep networks, notably pre-trained vision-language models (VLMs), is a formidable challenge. Current Class Activation Map (CAM) methods highlight regions revealing the model's decision-making basis but lack clear saliency maps and detailed interpretability. To bridge this gap, we propose DecomCAM, a novel decomposition-and-integration method that distills shared patterns from channel activation maps. Utilizing singular value decomposition, DecomCAM decomposes class-discriminative activation maps into orthogonal sub-saliency maps (OSSMs), which are then integrated together based on their contribution to the target concept. Extensive experiments on six benchmarks reveal that DecomCAM not only excels in locating accuracy but also achieves an optimizing balance between interpretability and computational efficiency. Further analysis unveils that OSSMs correlate with discernible object components, facilitating a granular understanding of the model's reasoning. This positions DecomCAM as a potential tool for fine-grained interpretation of advanced deep learning models. The code is avaible at https://github.com/CapricornGuang/DecomCAM.

Figures (9)

• Figure 1: Illustration of our DecomCAM method in action. We present the orthogonal sub-saliency maps produced by DecomCAM, which isolate distinct attributes within the images. This capability enables our method to handle complex concepts, e.g. "dog wearing a shirt", resulting in a comprehensive and clear saliency map that captures detailed aspects of the concept to be interpreted.
• Figure 2: DecomCAM overview. The proposed method operates through a streamlined two-stage process: a) In the decomposition stage, DecomCAM utilizes the gradient maps derived from predicted vectors and activation maps to create the class-discriminative activation maps, $\hat{S}^{g}$. Applying singular value decomposition to $\hat{S}^{g}$, we select the top-$Q$ Orthogonal Sub-Saliency Maps (OSSMs) that represent the top-$Q$ common activation patterns across different channels. These OSSMs indicate detailed visual features related to the target concept. b) In the integration stage, a Gaussian-blurred version of the input image and OSSMs are processed through the CNN to assess their respective impacts on the model's confidence. The difference in confidence levels informs the aggregation weights of OSSMs, which are then used to integrate OSSMs together to generate a clear and comprehensive saliency map.
• Figure 3: Visualization of saliency maps of different gradient-based CAMs. DecomCAM consistently outperforms other methods, producing saliency maps that are more focused on the target concept with less noise. The results are drawn from the CLIP-ResNet50x4 backbone. We generate the predicted bounding boxes following chen2022gscorecam.
• Figure 4: Visualization of saliency maps of different CAM methods. The ResNet50-CLIP is used. DecomCAM consistently outperforms other methods, showing a more focused and clear saliency map.
• Figure 5: Causal interpretation performance on PS-ImageNet. The graph illustrates how each method's interpretability metrics fluctuated with classification performance. Evaluation for various CAMs is performed using CLIP-ResNet50 backbone.