GFR-CAM: Gram-Schmidt Feature Reduction for Hierarchical Class Activation Maps

Kaveh Safavigerdini, Bahram Yaghooti, Amir Erfan Zareei Shams Abadi, Kannappan Palaniappan; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 744-753

Deep learning models have achieved remarkable success in computer vision tasks, yet their decision-making processes remain largely opaque, limiting their adoption especially in safety-critical applications. While Class Activation Maps (CAMs) have emerged as a prominent solution for visual explanation, existing methods suffer from a fundamental limitation: they produce single, consolidated explanations leading to "explanatory tunnel vision." Current CAM methods fail to capture the rich, multi-faceted reasoning that underlies model predictions, particularly in complex scenes with multiple objects or intricate visual relationships. We introduce the Gram-Schmidt Feature Reduction Class Activation Map (GFR-CAM), a novel gradient-free framework that overcomes this limitation through hierarchical feature decomposition that provides a more holistic view of the architecture's explanatory power. Unlike existing feature reduction methods that rely on Principal Component Analysis (PCA) and generate a single dominant explanation, GFR-CAM leverages Gram-Schmidt orthogonalization to systematically extract a sequence of orthogonal, information rich components from model feature maps. The subsequent orthogonal components are shown to be meaningful explanations - not mere noise, that decomposes single objects into semantic parts and systematically disentangles multi-object scenes to identify co-existing entities. We show that GFR-CAM on ResNet-50 and Swin Transformer architectures across ImageNet and PASCAL VOC datasets achieves competitive performance with state-of-the-art methods.