Research map / Label-efficient and robust learning / Recognition and long-tail
Modeling Saliency Dataset Bias
Matthias Kümmerer, Harneet Singh Khanuja, Matthias Bethge
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 changed | Where | Measured | Effect |
|---|---|---|---|
| train on another dataset then apply to unseen dataset (no adaptation) | saliency models · five saliency datasets | information gain: more than 40% information gain drop | Hurt |
| train on four other datasets instead of one | naive saliency model · five saliency datasets | inter-dataset gap: 58% of gap remains | No clear effect |
| adapt only the fewer-than-20 dataset bias parameters to the target dataset | bias-aware saliency model · five saliency datasets | generalization gap closed: 76% of the generalization gap closed | Improved |
| adapt bias parameters on only 50 images (ablation) | bias-aware saliency model · leave-one-dataset-out | information gain | Improved |
| add dataset-specific bias parameters to jointly trained model (ablation) | joint saliency model · five saliency datasets | information gain | Improved |
| adapt center bias and multiscale weights (ablation) | bias-aware saliency model · leave-one-dataset-out average | generalization gap closed | Improved |
| adapt priority scaling, blur size, center bias weight (ablation) | bias-aware saliency model · dataset average | generalization gap closed | Mixed |
| replace ensemble of backbones with CLIP+DINOv2, add multiscale extraction, make biases dataset dependent (ablation) | DeepGaze IIE · five-dataset combination | information gain | Improved |
| generalize via average-bias model vs naive model vs ensembling per-dataset models (ablation) | saliency models · generalization | information gain | No clear effect |
| model center bias as centered Gaussian instead of KDE (ablation) | center bias model · low-data adaptation | log-likelihood | Hurt |
| include large scales in multiscale feature extraction (ablation) | multiscale backbone · joint training | prediction performance | Improved |
| adapt bias parameters on SALICON | bias-aware saliency model · SALICON | information gain (bit/fix): -0.03 -> 0.26 | Improved |
Closest papers by meaning
- Saliency Prediction in the Deep Learning Era: Successes and Limitations (TPAMI 2019 · similarity 0.91)
- Understanding and Visualizing Deep Visual Saliency Models (CVPR 2019 · similarity 0.90)
- Saliency Benchmarking Made Easy: Separating Models, Maps and Metrics (ECCV 2018 · similarity 0.90)
- Personalized Saliency and Its Prediction (TPAMI 2018 · similarity 0.89)
- Revisiting Video Saliency Prediction in the Deep Learning Era (TPAMI 2019 · similarity 0.89)
- Advancing Saliency Ranking with Human Fixations: Dataset Models and Benchmarks (CVPR 2024 · similarity 0.87)
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.
