Research map / Label-efficient and robust learning / Recognition and long-tail

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

CVPR 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Self-explaining and visual explanation networks · 14 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Finer-CAM, which turns any gradient- or score-based CAM into a comparison between the target class logit and a similar class logit, highlighting fine-grained discriminative regions without extra explanation complexity.

Problem

Standard CAM methods highlight all regions that contribute to a target class logit, including regions predictive of visually similar classes, so they fail to localize discriminative details in fine-grained classification.

Approach

Replace the target-class logit in CAM with the logit difference between the target and similar classes (y_c - γ·y_d), so features shared with similar classes are suppressed; works for Grad-CAM, Score-CAM, and zero-shot CLIP text prompts, with multi-reference aggregation and a relative confidence drop metric.

Key results

• On Birds-525 with Grad-CAM, Finer-CAM raises RD@0.05 from 0.174 to 0.192 and RD@0.1 from 0.245 to 0.260.
• On CUB-200 pointing game, Layer-CAM localization improves from 0.625 to 0.682 with Finer-CAM.
• Masking top 5% pixels yields a larger relative confidence drop than baselines across Grad-CAM, Layer-CAM, and Score-CAM; deletion AUC differences are negligible.

Measured effects reported in the paper

Surprising findings

• Standard deletion AUC shows only negligible differences between baselines and Finer-CAM, so the authors rely on a new relative drop metric instead.
• Direct subtraction of similar-class saliency maps produces noisy results because of the ReLU, unlike Finer-CAM's logit-level difference.

Limitations

• Relies on the choice of similar reference classes (top-3 similar classes via classifier weights or predicted rank), and effectiveness depends on that choice.
• Deletion AUC gains are negligible; the advantage is shown mainly via the proposed relative drop metric and pointing game.

Experimental setup

Datasets: Birds-525 CUB-200 Cars Aircraft FishVista

Benchmarks: Birds-525 CUB-200 Cars

Metrics: deletion AUC relative confidence drop (RD@0.05, RD@0.1) energy-based pointing game localization

Baselines: Grad-CAM Layer-CAM Score-CAM

Builds on: Grad-CAM Score-CAM Layer-CAM CLIP

Closest papers by meaning

• DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging (NeurIPS 2026 · similarity 0.91)
• Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness Explanations (AAAI 2024 · similarity 0.88)
• CAPE: CAM as a Probabilistic Ensemble for Enhanced DNN Interpretation (CVPR 2024 · similarity 0.88)
• Prompt-CAM: Making Vision Transformers Interpretable for Fine-Grained Analysis (CVPR 2025 · similarity 0.87)
• Towards Fine-Grained Interpretability: Counterfactual Explanations for Misclassification with Saliency Partition (CVPR 2025 · similarity 0.86)
• VisMin: Visual Minimal-Change Understanding (NeurIPS 2024 · similarity 0.86)