Research map / Detection and segmentation / CNN features
OAT: Object-Level Attention Transformer for Gaze Scanpath Prediction
Yini Fang, Jingling Yu, Haozheng Zhang, Ralf van der Lans, Bertram E Shi
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 |
|---|---|---|---|
| replace end-to-end learnable PE with distance-based PE (ablation) | OAT · Amazon book cover | Overall behaviour difference: 0.088 -> 0.074 | Improved |
| add relative positional encoding on top of DPE (ablation) | OAT · Amazon book cover | Overall behaviour difference: 0.074 -> 0.088 | Hurt |
| add distance-based PE (ablation) | OAT · Yogurt/wine (collected) | Overall behaviour difference: 0.119 -> 0.031 | Improved |
| add object attention module (ablation) | OAT · Yogurt/wine (collected) | Overall behaviour difference: 0.169 -> 0.031 | Improved |
| add both DPE and OA (ablation) | OAT · Yogurt/wine (collected) | Sequence Score: 0.224 -> 0.346 | Improved |
| tokenize at object level instead of sub-object grid (12x28 patches) (ablation) | OAT · Amazon book cover | Overall behaviour difference: 1.14 -> 0.07 | Improved |
| coarser grid of 3x7 (patches of four objects) vs object-level 6x14 (ablation) | OAT · Amazon book cover (mapped to 3x7) | Overall behaviour difference: 0.05 -> 0.01 | No clear effect |
| model scanpaths at object level with autoregressive decoding vs Gazeformer | OAT · Amazon book cover | Sequence Score: 0.098 -> 0.950 | Improved |
| OAT vs Gazeformer on FED in collected dataset | OAT · Yogurt/wine (collected) | FED: 7.69 -> 7.19 | Improved |
| OAT on MultiMatch direction score in collected dataset vs IRL | OAT · Yogurt/wine (collected) | MultiMatch direction: 0.603 vs 0.645 (IRL) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Gazeformer — Gazeformer: Scalable, Effective and Fast Prediction of Goal-Directed Human Attention (CVPR 2023) · better on 6 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unifying Top-down and Bottom-up Scanpath Prediction Using Transformers (CVPR 2024 · similarity 0.90)
- Gazeformer: Scalable, Effective and Fast Prediction of Goal-Directed Human Attention (CVPR 2023 · similarity 0.89)
- TransGOP: Transformer-Based Gaze Object Prediction (AAAI 2024 · similarity 0.87)
- Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders (CVPR 2025 · similarity 0.87)
- SeqRank: Sequential Ranking of Salient Objects (AAAI 2024 · similarity 0.86)
- EyEar: Learning Audio Synchronized Human Gaze Trajectory Based on Physics-Informed Dynamics (AAAI 2025 · similarity 0.86)
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.
