Debiasing Central Fixation Confounds Reveals a Peripheral “Sweet Spot” for Human-like Scanpaths in Hard-Attention Vision
Pengcheng Pan, Yonekura Shogo, Yasuo Kuniyoshi · 2026
Read Debiasing Central Fixation Confounds Reveals a Peripheral “Sweet Spot” for Human-like Scanpaths in Hard-Attention Vision by Pengcheng Pan, Yonekura Shogo, Yasuo Kuniyoshi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-14834.txt.
Description
Pengcheng Pan, Yonekura Shogo, and Yasuo Kuniyosh show that scanpath metrics for hard-attention vision can be fooled by centre bias. On Gaze-CIFAR-10, a trivial centre-fixation baseline scores surprisingly well, nearing many learned policies, because object-centric datasets reward looking at the middle regardless of behaviour.
Debiasing that confound reveals a peripheral sweet spot for human-like scanpaths, where foveal patch size and peripheral context are swept under constrained vision. Genuine gaze alignment separates from mere central tendency only once the baseline is stripped out. Vision models gain a fairer test of how they actually look.
Catalog record
| CREATOR | Pengcheng Pan, Yonekura Shogo, Yasuo Kuniyoshi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-14834