AI Paper Atlas

Research map / Architectures and efficiency / CNN design

Winsor-CAM: Human-Tunable Visual Explanations From Deep Networks via Layer-Wise Winsorization

Casey Wall, L. Wang, Rodrigue Rizk, KC Santosh

TPAMI 2026 · Architectures and efficiency › CNN design · approach: Biologically inspired CNN explanation

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add layer-wise Winsorized aggregation over all conv layers instead of final-layer Grad-CAMGrad-CAM (DenseNet121) · PASCAL VOC 2012 subset (correct)IoU: 0.390 -> 0.468Improved
add Winsorization instead of naive mean aggregation of layer mapsnaive layer averaging (DenseNet121) · PASCAL VOC 2012 subset (correct)IoU: 0.428 -> 0.468Improved
add Winsorization instead of final-layer Grad-CAMGrad-CAM (DenseNet121) · PASCAL VOC 2012 subset (correct)CoM distance: 0.074 -> 0.059Improved
add Winsorization instead of final-layer Grad-CAMGrad-CAM (DenseNet121) · PASCAL VOC 2012 subset (correct)Deletion AUC: 0.242 -> 0.197Improved
add Winsorization instead of final-layer Grad-CAMGrad-CAM (VGG16) · PASCAL VOC 2012 subset (correct)Insertion AUC: 0.900 -> 0.873Hurt
add Winsorization instead of final-layer Grad-CAMGrad-CAM (EfficientNet-B0) · PASCAL VOC 2012 subset (correct)IoU: 0.398 -> 0.391Hurt
use bilinear interpolation instead of nearest-neighbor (ablation)Winsor-CAM (DenseNet121, mean) · PASCAL VOC 2012 subset (correct)IoU: 0.456 -> 0.468Improved
use mean aggregation instead of max aggregation (ablation)Winsor-CAM (DenseNet121, bilinear) · PASCAL VOC 2012 subset (correct)IoU: 0.453 -> 0.468Improved
use max aggregation instead of mean aggregation (ablation)Winsor-CAM (VGG16, bilinear) · PASCAL VOC 2012 subset (correct)Insertion AUC: 0.873 -> 0.905Improved
increase fixed percentile p from 0 to 50 (ablation)Winsor-CAM (DenseNet121) · PASCAL VOC 2012 subset (correct)IoU: 0.442 -> 0.445No clear effect
increase fixed percentile p from 0 to 90 (ablation)Winsor-CAM (DenseNet121) · PASCAL VOC 2012 subset (correct)Deletion AUC: 0.208 -> 0.215Hurt
tune p per image (best over p) instead of one fixed p (ablation)Winsor-CAM (DenseNet121) · PASCAL VOC 2012 subset (correct)IoU: 0.445 -> 0.468Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy