Reading Decisions from Gaze Direction during Graphics Turing Test of Gait Animation

TL;DR

This paper examines how gaze behavior during observation of natural versus MP-generated gait relates to perceptual decisions in a graphics Turing test. By combining wearable eye-tracking with pose-keypoint tracking via DeepLabCut and employing Bayesian logistic regression and mutual information analyses, the authors quantify how gaze features correspond to decisions, task context, and accuracy. They find that gaze information aligns more with the participants' reported decisions than with whether those decisions were correct, with the last fixation emerging as the strongest predictor; the gaze cascade result is inconclusive. The work suggests eye-tracking can meaningfully augment binary decision data in perceptual movement studies and informs design considerations for future gaze-informed experiments.

Abstract

We investigated gaze direction during movement observation. The eye movement data were collected during an experiment, in which different models of movement production (based on movement primitives, MPs) were compared in a two alternatives forced choice task (2AFC). Participants observed side-by-side presentation of two naturalistic 3D-rendered human movement videos, where one video was based on motion captured gait sequence, the other one was generated by recombining the machine-learned MPs to approximate the same movement. The task was to discriminate between these movements while their eye movements were recorded. We are complementing previous binary decision data analyses with eye tracking data. Here, we are investigating the role of gaze direction during task execution. We computed the shared information between gaze features and decisions of the participants, and between gaze features and correct answers. We found that eye movements reflect the decision of participants during the 2AFC task, but not the correct answer. This result is important for future experiments, which should take advantage of eye tracking to complement binary decision data.

Figures (7)

• Figure 1: Figure reprint with permission of knoppPredictingPerceivedNaturalness2018: Illustration of experimental procedure. Each trial begun with a fixation period of 0.75s. Then, participants watched simultaneous replays of natural and generated movements for 3.5s. After the presentation the participants were asked "On which side did you perceive the more natural movement?" and responded using the arrow keys of a standard computer keyboard.
• Figure 2: Training labels are marked with "+", predictions by the model with "$\cdot$". Four green squares marked the display corners. Predictions with likelihood $\leq p_{cutoff}$, are marked with "$\times$" (see occluded knee of left avatar).
• Figure 3: Upper pane: Gaze trajectory of one exemplary trial. The four green dots in the corners show the monitor corner marks (cf. fig. \ref{['fig:labels']}), and the fixation cross where the dashed red lines meet. The complete gaze trajectory is dashed blue, with small blue points marking the gaze direction in each frame. Each segment, where participants gaze follows the avatar is plotted as orange, green, red and purple lines. Below: Speed (angles / second) for the trajectory in blue. Speeds above $100^\circ/s$ segment the trajectory, leaving the segments (color coded as above) of single avatar gazing. See text for more detailed description.
• Figure 4: Histogram of posterior samples for model parameter $\beta_{\mathrm{last fixation}}$ to predict outcome variable response (decision). Violet line is the kernel-density estimation of the posterior distribution, red line is the normal distribution with parameters estimated from the samples. The normal is thus a very good approximation to the posterior. The green line shows the prior: $\mathcal{N}(0, 1)$, which seems almost uniform in comparison to the posterior. Thus, the data overrules the prior.
• Figure 5: Gaze speed distribution. Bars show normalized bin counts, lines a Gaussian kernel density estimation in corresponding color. Color codes participant ID.