Finer-CAM: Spotting the Difference Reveals Finer Details for Visual Explanation

• DOI: 10.48550/arxiv.2501.11309
• Publication Date: 2025-01-20
• Authors: Ziheng Zhang; Jianyang Gu; Arpita Saha Chowd...; Zheda Mai; David Carlyn; Tanya Berger‐Wolf; Yu Su; Weilun Chao
• Keywords: Spotting

Class activation map (CAM) has been widely used to highlight image regions that contribute class predictions. Despite its simplicity and computational efficiency, CAM often struggles identify discriminative distinguish visually similar fine-grained classes. Prior efforts address this limitation by introducing more sophisticated explanation processes, but at the cost of extra complexity. In paper, we propose Finer-CAM, a method retains CAM's efficiency while achieving precise localization regions. Our key insight is deficiency lies not in "how" it explains, "what" explains}. Specifically, previous methods attempt all cues contributing target class's logit value, which inadvertently also activates predictive By explicitly comparing with classes spotting their differences, Finer-CAM suppresses features shared other emphasizes unique, details class. easy implement, compatible various methods, can be extended multi-modal models for accurate specific concepts. Additionally, allows adjustable comparison strength, enabling users selectively coarse object contours or fine details. Quantitatively, show masking out top 5% activated pixels results larger relative confidence drop compared baselines. The source code demo are available https://github.com/Imageomics/Finer-CAM.