Class Activation Mapping in Explainable Computer Vision: A Method-Centric Survey

Forum topic · 小凯 · 2026-08-14

Summary

This paper surveys 57 method-centric publications on Class Activation Mapping (CAM), one of the most widely used visual explanation families in explainable artificial intelligence. Since the original CAM formulation in 2016 for global-average-pooling CNN classifiers, CAM-style methods have expanded to include gradient-based post-hoc explanations, gradient-free scoring and ablation techniques, high-resolution upsampling, weakly supervised localization and segmentation, transformer token attribution, causal and debiasing approaches, and foundation-model-era methods using CLIP, DINO, SAM, and feature-distribution comparisons. The authors develop a taxonomy that separates methods by attribution mechanism, architectural dependence, and evaluation objective, covering gradient-based CAM, recent and hybrid CAM-style approaches, and model- or architecture-aware methods. A key trend identified is a shift from explaining single-class scores in single-resolution CNN layers toward comparative, multi-layer, probabilistic, token-aware, and foundation-model-aware explanations. The survey also highlights that evaluation remains fragmented, with faithfulness, localization, robustness, computational cost, and human trust measured under inconsistent protocols, pointing to open research gaps.

Paper Overview

Field: Computer Vision (Explainable AI)
Authors: AmirHossein Eshghi, Hamid Saadatfar, Seyyed Ali Hoseini
Release Date: 2026-08-13
arXiv: 2508.03414

Abstract

Class Activation Mapping (CAM) is one of the most widely used visual explanation families in explainable artificial intelligence. Its purpose is intuitive: to convert internal model evidence into heatmaps that highlight the image regions, convolutional channels, tokens, or patches supporting a target class or concept. Since the first CAM formulation in 2016, the field has moved well beyond global-average-pooling CNN classifiers. CAM-style methods now include gradient-based post-hoc explanations, gradient-free scoring and ablation methods, high-resolution upsampling, weakly supervised localization and segmentation, transformer token attribution, causal and debiasing approaches, and foundation-model-era methods using CLIP, DINO, SAM, or feature-distribution comparisons.

This survey synthesizes 57 rigorously selected method-centric publications published since 2016. The authors develop a taxonomy that separates methods by attribution mechanism, architectural dependence, and evaluation objective. The survey then reviews gradient-based CAM, recent and hybrid CAM-style methods, and model- or architecture-aware methods.

Key Points

Scope: A method-centric synthesis of 57 papers on CAM-based visual explanations published since 2016.
Methodological breadth: Coverage spans gradient-based post-hoc explanations, gradient-free scoring/ablation, high-resolution upsampling, weakly supervised localization and segmentation, transformer token attribution, and causal/debiasing approaches.
Modern era: Foundation-model methods leveraging CLIP, DINO, SAM, and feature-distribution comparisons are explicitly included.
Taxonomy: Methods are organized along three axes: attribution mechanism, architectural dependence, and evaluation objective.
Major trend: The field is shifting from explaining single-class scores in single-resolution CNN layers toward comparative, multi-layer, probabilistic, token-aware, and foundation-model-aware explanations.
Evaluation gap: Assessment remains fragmented; faithfulness, localization, robustness, computational cost, and human trust are typically measured under different and incompatible protocols.
Contribution emphasis: For each method, the survey highlights both its contribution and the gaps it leaves, as well as how subsequent methods attempt to fill those gaps.

*Auto-collected on 2026-08-14.*