Predicting Selection Intention in Real-Time with Bayesian-based ML Model in Unimodal Gaze Interaction

TL;DR

A Bayesian-based machine learning model is presented to predict user selection intention in real-time using only gaze data and it is found that the selection intention inferred by the model enables more comfortable and accurate interactions compared to traditional techniques.

Abstract

Eye gaze is considered a promising interaction modality in extende reality (XR) environments. However, determining selection intention from gaze data often requires additional manual selection techniques. We present a Bayesian-based machine learning (ML) model to predict user selection intention in real-time using only gaze data. Our model uses a Bayesian approach to transform gaze data into selection probabilities, which are then fed into an ML model to discriminate selection intentions. In Study 1, our model achieved real-time inference with an accuracy of 0.97 and an F1 score of 0.96. In Study 2, we found that the selection intention inferred by our model enables more comfortable and accurate interactions compared to traditional techniques.

Figures (5)

• Figure 1: We configured a total of four types of target configurations by combining two factors: target size (Large or Small) and target density (Wide or Dense).
• Figure 2: We categorized acquisition complexity into two levels based on the difficulty of identifying the target. In the Simple acquisition task (left), a red target object is placed among blue distractors, and the task is to click on the target. In the Complex acquisition task (right), each distractor has a random word written on it, and the task is to find and click the target object that has a word belonging to the same category as the word written on the starting object.
• Figure 3: Time-to-Completion (TTC) (Left) and Error Rate (Right) based on target configuration and input modality. Statistically significant differences identified through post-hoc tests are indicated with connecting lines, where *, **, and *** represent p-values less than 0.05, 0.01, and 0.001, respectively.
• Figure 4: Overall NASA-TLX score (Left) and physical demand (Right) based on target configuration and input modality. Statistically significant differences identified through post-hoc tests are indicated with connecting lines, where *, **, and *** represent p-values less than 0.05, 0.01, and 0.001, respectively.
• Figure 5: SUS survey score (Left) and user preference rank score collected through interviews (Right) based on target configuration and input modality. Statistically significant differences identified through post-hoc tests are indicated with connecting lines, where *, **, and *** represent p-values less than 0.05, 0.01, and 0.001, respectively.