Research map / Label-efficient and robust learning / Recognition and long-tail
DiffCAM: Data-Driven Saliency Maps by Capturing Feature Differences
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 DiffCAM, a CAM-style saliency method that picks the feature-importance direction maximizing the separation between a target example and a reference data distribution, rather than following decision-boundary gradients.
Problem
Existing CAM-based saliency methods estimate feature importance from decision-boundary gradients or prediction-dependent scores, which can be counter-intuitive for non-representative examples and do not capture the actual reference-class data distribution.
Approach
Define feature importance as the direction w maximizing the ratio of squared projected distance between the target and the reference mean to the projected within-reference variance (Fisher-style closed form w* = Sv^-1 (z - mu)), with flexible reference sets enabling "why is", "why not", and intra-class variance explanations, without using prediction outputs.
Key results
- On ImageNet object localization, DiffCAM reaches IoU 0.534 and AUPRC 0.741 with ResNet-50 and IoU 0.583 / AUPRC 0.794 with Inception-v3, matching or tying the best CAM baselines.
- On CUB-200-2011 counterfactual explanation, DiffCAM achieves recall 0.17 and precision 0.83 for random counterfactual classes and recall 0.17 with precision 0.61 for similar classes, outperforming GALORE overall.
- On RSNA pneumonia localization with Inception-v3, DiffCAM reaches AUPRC 0.492 versus 0.467 for GradCAM and 0.328 for ScoreCAM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffCAM instead of LayerCAM for saliency-map object localization | ResNet-50 · ImageNet | AUPRC: 0.738 -> 0.741 | Helped |
| Use DiffCAM instead of LayerCAM for saliency-map object localization | Inception-v3 · ImageNet | IoU: 0.583 -> 0.583 | No clear effect |
| Use DiffCAM instead of LayerCAM for saliency-map object localization | Inception-v3 · ImageNet | AUPRC: 0.792 -> 0.794 | Helped |
| Use DiffCAM instead of GALORE for counterfactual saliency maps | ResNet-50 · CUB-200-2011, random counterfactual | R: 0.09 -> 0.15 | Helped |
| Use DiffCAM instead of CounteRGAN for counterfactual saliency maps | ResNet-50 · CUB-200-2011, random counterfactual | PIoU: 0.20 -> 0.18 | Hurt |
| Use DiffCAM instead of GALORE for counterfactual saliency maps | ResNet-50 · CUB-200-2011, similar counterfactual | PIoU: 0.15 -> 0.20 | Helped |
| Use DiffCAM instead of GradCAM for abnormality localization | Inception-v3 · RSNA Pneumonia Detection Challenge | AUPRC: 0.467 -> 0.492 | Helped |
| Use DiffCAM instead of GradCAM for abnormality localization | Inception-v3 · SIIM-ACR Pneumothorax Segmentation | AUPRC: 0.076 -> 0.093 | Helped |
Surprising findings
- On the SIIM pneumothorax task, all methods yield low AUPRC (about 0.06 to 0.09), though DiffCAM is still highest at 0.093.
- The paper reports that gradient-based and CAM-based methods can prioritize features counter-intuitively for non-representative examples, e.g., attributing a failure to writing score even when the target's writing score is high.
Limitations
- Assumes a linear classifier is attached to the final deep feature layer, so generality to other architectures is left to future work.
- Results on the SIIM pneumothorax task are low for all methods, indicating limited localization accuracy on hard medical data.
Experimental setup
Datasets: MNIST ImageNet CUB-200-2011 RSNA Pneumonia Detection Challenge SIIM-ACR Pneumothorax Segmentation
Benchmarks: ImageNet CUB-200-2011 RSNA Pneumonia Detection Challenge SIIM-ACR Pneumothorax Segmentation MNIST
Metrics: ROAD faithfulness IoU AUPRC precision recall PIoU F1
Baselines: GradCAM ScoreCAM LayerCAM XGradCAM Integrated Gradients Guided Backpropagation GALORE CounteRGAN Goyal et al. counterfactual
Builds on: CAM GradCAM ScoreCAM GALORE
Closest papers by meaning
- DD-CAM: Minimal Sufficient Explanations for Vision Models Using Delta Debugging (NeurIPS 2026 · similarity 0.88)
- CAPE: CAM as a Probabilistic Ensemble for Enhanced DNN Interpretation (CVPR 2024 · similarity 0.86)
- Explainable Saliency: Articulating Reasoning with Contextual Prioritization (CVPR 2025 · similarity 0.86)
- Explaining in Diffusion: Explaining a Classifier with Diffusion Semantics (CVPR 2025 · similarity 0.85)
- Finer-CAM: Spotting the Difference Reveals Finer Details for Visual Explanation (CVPR 2025 · similarity 0.84)
- Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness Explanations (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.