Real-time head-based deep-learning model for gaze probability regions in collaborative VR

Abstract

Figure 1: a) The head rotational velocity direction (θ ) and magnitude (ρ) are extracted during a VR session. b) Probability density functions are extracted from eye-gaze distributions that correspond to the head rotational velocity and are converted into a series of percentile-based contours (η). c) Our real-time model uses the three parameters (θ ,ρ,η) to provide a novel representation of visual attention for VR collaboration or interaction.

Fovea Prediction Model in VR

“…The model employed is an MLP neural network tailored for gaze estimation. We introduced the head's pitch parameter compared to Bovo's study [2] referred to as "latitude"(φ ). The latitude, as a factor in our predictive model, played a crucial role in improving the accuracy of gaze prediction.…”

“…It relies exclusively on head movements, making it suitable for real-time deployment on affordable HMDs. Our choice of the multi-perceptron (MLP) neural network (NN) is informed by its successful results in Bovo's work [2]. Our MLP architecture is a fully connected layer designed to learn the central fixation points and the outlines of probable fixation areas.…”

Classifying Head Movements to Separate Head-Gaze and Head Gestures as Distinct Modes of Input

“…Head movements play a signifcant role in interaction as a means of expressing intent and eliciting emotions. Previous work has demonstrated the potential of machine learning models for detecting and classifying head gestures in a wide range of application scenarios [4,15,40,58]. For instance, Morency et al leveraged visual features (e.g., head velocities or eyegaze estimates) to propose an SVM model to classify head gestures as feedback nods and headshakes during interaction [35].…”