AI Paper Atlas

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

Modeling Saliency Dataset Bias

Matthias Kümmerer, Harneet Singh Khanuja, Matthias Bethge

ICCV 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning · 4 citations (OpenAlex)

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
train on another dataset then apply to unseen dataset (no adaptation)saliency models · five saliency datasetsinformation gain: more than 40% information gain dropHurt
train on four other datasets instead of onenaive saliency model · five saliency datasetsinter-dataset gap: 58% of gap remainsNo clear effect
adapt only the fewer-than-20 dataset bias parameters to the target datasetbias-aware saliency model · five saliency datasetsgeneralization gap closed: 76% of the generalization gap closedImproved
adapt bias parameters on only 50 images (ablation)bias-aware saliency model · leave-one-dataset-outinformation gainImproved
add dataset-specific bias parameters to jointly trained model (ablation)joint saliency model · five saliency datasetsinformation gainImproved
adapt center bias and multiscale weights (ablation)bias-aware saliency model · leave-one-dataset-out averagegeneralization gap closedImproved
adapt priority scaling, blur size, center bias weight (ablation)bias-aware saliency model · dataset averagegeneralization gap closedMixed
replace ensemble of backbones with CLIP+DINOv2, add multiscale extraction, make biases dataset dependent (ablation)DeepGaze IIE · five-dataset combinationinformation gainImproved
generalize via average-bias model vs naive model vs ensembling per-dataset models (ablation)saliency models · generalizationinformation gainNo clear effect
model center bias as centered Gaussian instead of KDE (ablation)center bias model · low-data adaptationlog-likelihoodHurt
include large scales in multiscale feature extraction (ablation)multiscale backbone · joint trainingprediction performanceImproved
adapt bias parameters on SALICONbias-aware saliency model · SALICONinformation gain (bit/fix): -0.03 -> 0.26Improved

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