The Eye-Head Mover Spectrum: Modelling Individual and Population Head Movement Tendencies in Virtual Reality

TL;DR

The paper defines the eye-head mover spectrum as a continuous dimension of individual variation in VR gaze control, quantifying how head contribution to gaze shifts scales with target eccentricity. It introduces a soft-hinge parametric model fitted to per-participant data from a large 360° VR free-viewing dataset, then derives a population distribution via functional PCA. A controlled user study confirms cross-task stability while showing context-driven shifts in the distribution shape, indicating that tendencies are largely robust but modulated by task demands. The work demonstrates high relevance for VR system design, including adaptive foveated rendering and viewport alignment, and provides a principled framework for incorporating individual coordination differences into immersive experiences. Overall, the eye-head mover spectrum offers a rigorous, learnable representation of a previously underexplored source of individual variability with clear practical implications for accessibility and personalization in VR/AR.

Abstract

People differ in how much they move their head versus their eyes when shifting gaze, yet such tendencies remain largely unexplored in HCI. We introduce head movement tendencies as a fundamental dimension of individual difference in VR and provide a quantitative account of their population-level distribution. Using a 360° video free-viewing dataset (N=87), we model head contributions to gaze shifts with a hinge-based parametric function, revealing a spectrum of strategies from eye-movers to head-movers. We then conduct a user study (N=28) combining 360° video viewing with a short controlled task using gaze targets. While parameter values differ across tasks, individuals show partial alignment in their relative positions within the population, indicating that tendencies are meaningful but shaped by context. Our findings establish head movement tendencies as an important concept for VR and highlight implications for adaptive systems such as foveated rendering, viewport alignment, and multi-user experience design.

Figures (12)

• Figure 1: Example fits of three model formulations to one participant (P05). Each plot shows horizontal head contribution ($\Delta$Head) against target eccentricity, with grey dots as data points. (a) Linear + EOR baseline assumes a fixed eye-only range followed by a linear increase. (b) Hinge model introduces a sharp breakpoint between eye- and head-dominant behaviour. (c) Soft hinge extends this with a smooth logistic transition, capturing both early and gradual head involvement.
• Figure 2: Comparison of model fits for gaze shift data. Boxplots show the distribution of R² (left) and RMSE (right) across participants for three models: Linear baseline, two-parameter hinge, and three-parameter soft hinge. Boxes represent the interquartile range (IQR) with medians; black dots indicate group means. Significance bars show pairwise differences between models ($*p < .05$, $**p < .01$, $***p < .001$).
• Figure 3: Main spectrum of variation in head contribution revealed by fPCA across participants in the D-SAV360 dataset. fPCA identifies the dominant ways behavioural curves vary across participants. (a) Reconstruction of the first fPCA component (91.1% variance explained). The average curve is shown in black, with dashed curves indicating ±2 standard deviations along this component. (b) Participants at the two ends of the same spectrum. Left: 5 participants with lower head contribution. Right: 5 participants with higher head contribution. Thin lines represent individual participants; the thick black line is the median curve within each group.
• Figure 4: Secondary spectrum of variation identified by fPCA (D-SAV360 dataset). The second fPCA component explains 7.8% of the variance in the data. The average curve is shown in black, with dashed curves indicating ±2 standard deviations along this component. This secondary spectrum reflects subtler differences in head movement that go beyond the main continuum.
• Figure 5: Participant distribution on the main spectrum of variation identified by fPCA (D-SAV360 dataset). The x-axis shows each participant’s score on the first functional principal component (fPCA), which captures the continuum from lower to higher head contribution in gaze shifts. The smooth curve shows the overall distribution, dots mark individual participants, and the vertical lines indicate the 25th percentile (Q1), median, and 75th percentile (Q3). The numbers are the participants' IDs which mapped to the curves in \ref{['fig:fpca_360_pc1']}(b).