Debiasing Central Fixation Confounds Reveals a Peripheral Sweet Spot for Human-like Scanpaths in Hard-Attention Vision

Jul 16, 2026

Upload by Finn

Human eye movements balance high-resolution foveal sampling with informative peripheral context, yet judging whether artificial scanpaths are truly human-like is hindered by object-centric benchmarks that exhibit strong center bias. Using Gaze-CIFAR-10, a center-fixation baseline attains unexpectedly high similarity under common metrics, masking strategy alignment. A new GCS score debiases and accounts for movement, enabling center-debiased evaluation. Under hard-attention constraints, the best human-like resemblance emerges in a restricted mid-range regime: moderate foveal patches plus peripheral context outperform both narrower and broader alternatives.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages7

Reading Time18 min

Abstract

Introduction

Center bias as a confound in scanpath evaluation

Active vision, perceptual constraints, and strategy regimes

Debiasing Central Fixation Confounds Reveals a Peripheral “Sweet Spot” for Human-like Scanpaths in Hard-Attention Vision Pengcheng Pan 1 , Yonekura Shogo 1 & Yasuo Kuniyoshi 1 1Department of Mechano-Informatics, The University of Tokyo arXiv :2602 . 14834v2 [ cs .CV] 13 Jul 2026 Abstract Human eye movements in visual recognition reflect a balance between foveal sampling and peripheral context, but evaluating whether artificial scanpaths are "human-like" is difficult on object-centric datasets with strong center bias. Using Gaze-CIFAR-10, we show that a trivial center-fixation baseline achieves surprisingly strong scores under common scanpath metrics, blurring the distinction between behavioral alignment and central tendency. We introduce GCS (Gaze Consistency Score), a practical center-debiased and movement-aware score that normalizes against human and corner references, subtracts the center baseline, and adds a small movement-similarity term. Applying GCS to a hard-attention classifier under varied fovea– periphery constraints identifies a restricted mid-range regime: a moderate foveal patch with peripheral context yields stronger center-debiased alignment than either narrower or broader alternatives. This regime is not identified by accuracy alone; the highest-accuracy setting differs from the best-GCS setting. These results highlight the need for bias-aware scanpath evaluation and suggest that, on Gaze-CIFAR-10 and under this hardattention setting, perceptual constraints shape when task-trained policies appear relatively human-like. Keywords: active perception; hard attention; scanpath similarity; center bias; gaze metrics; peripheral vision Introduction Human visual perception is an active process. Rather than processing the entire visual field uniformly, humans rely on a foveated visual system and sequential eye movements to gather task-relevant information. Classic work by Yarbus, 1967 demonstrated that eye movement patterns are systematically shaped by task demands, suggesting that scanpaths can provide a window into underlying cognitive strategies. In recent years, this idea has motivated the development of hard-attention models in machine vision, which explicitly select where to look at each step while performing tasks such as image classification or visual search (Ba et al., 2014; Mnihet al., 2014; Pan et al., 2026) . These models offer a promising computational framework for studying perception under resource constraints. However, an open question remains: when do the scanpaths produced by such models trained purely on the task meaningfully resemble human eye movements? A major challenge in answering this question lies in how scanpaths are evaluated. A wide range of metrics have been proposed to compare fixation sequences, made available on FixaTons tools (Zanca et al., 2018), including Dynamic Time Warping (“Dynamic Time Warping,” 2007), ScanMatch (Cristino et al., 2010), Normalized Scanpath Saliency (Peterset al., 2005), and AUC-based measures originally developed for saliency evaluation (Judd et al., 2012) . These metrics capture complementary aspects of scanpaths, such as temporal alignment, spatial overlap, and distributional similarity. However, a long-standing finding in human eye-movement research complicates the interpretation: human fixations are strongly biased toward the image center. This central fixation bias has been robustly documented across free viewing and task-driven settings (Bindemann, 2010; Tatler, 2007; Tatler et al., 2011) . Proposed explanations include photographer bias, experimental framing, oculomotor constraints, and learned expectations about where informative content is likely to appear (Smith & Mital, 2013) . Crucially, center bias is not merely noise, it is a systematic property of many gaze datasets. Center bias as a confound in scanpath evaluation. When center bias is strong, scanpath similarity metrics can be dominated by marginal spatial distributions

FAQs

Why is evaluating “human-like” scanpaths difficult on object-centric vision datasets?

Because strong center bias can make similarity metrics reflect spatial tendencies rather than task-driven gaze strategy, allowing trivial behaviors to appear human-like.

What problem does GCS (Gaze Consistency Score) address?

It normalizes and debiases scanpath comparison by reducing influence from central tendencies and chance-like references, while also incorporating movement awareness.

What scanpath regime produces the most human-like alignment under hard-attention constraints?

A restricted mid-range setting works best: a moderate foveal patch combined with peripheral context yields stronger center-debiased alignment than both narrower and broader alternatives.

Jul 16, 2026

Upload by Finn