An oculomotor plant mathematical model (OPMM) employs physical and neurological characteristics of human visual system to define its dynamics. One of its most prominent applications in modern eye-tracking pipelines was hypothesized to be latency reduction via the means of eye movement prediction. However, this use case was only explored with OPMMs originally designed for saccade simulation. Such models typically relied on the neural pulse control being estimated from intended saccade amplitude - a property that becomes fully observed only after a saccade already ended, which greatly limits the model’s prediction capabilities. We present the first OPMM designed with the prediction task in mind. We draw our inspiration from a "peak velocity - amplitude" main sequence relationship and propose to use saccade’s peak velocity for neural pulse estimation. We additionally extend the prior work by evaluating the proposed model on the largest to date pool of 322 subjects against the naive zero displacement baseline and a long short-term memory (LSTM) neural network.
17 May 2024 in Proceedings of the ACM on Computer Graphics and Interactive Techniques
The practical value of oculomotor plant mathematical models (OPMMs) has been demonstrated across various domains, including biometrics and eye movement prediction. To further enhance their utilization, new optimization approaches are commonly developed and introduced within the research community. In this study, we demonstrate a new integration of a previously developed per-subject optimization procedure for an Enderle OPMM and introduce methods to evaluate the reliability of OPMM parameters using the intraclass correlation coefficient (ICC) and Kendall's Coefficient of Concordance (KCC). We evaluated two per-subject OPMM models, Bahill and Enderle, using the 'GazeBase' eye movement dataset. The models differed in accuracy, expressed as the per-sample error in degrees of visual angle, based on the differences between the actual and predicted saccade trajectories. We found that some of the parameter estimates for both models were quite unreliable. We discuss the importance of addressing low-reliability issues and suggest methods to modify the models to enhance reliability and performance.
This manuscript demonstrates an improved model-based approach for synthetic degradation of pre-recorded eye movement signals. Recordings from a high-quality eye tracking sensor are transformed to make their eye tracking signal quality resemble ones captured on a lower-quality target device. The proposed model improves the realism of the degraded signals versus prior approaches by introducing a mechanism for degrading spatial accuracy and temporal precision. Specifically, a percentile-matching technique is developed for mimicking the relative distributional structure of the target data signal quality characteristics. The model is demonstrated to improve realism on a per-feature and per-recording basis using data from EyeLink 1000 and SMI eye tracker embedded within a virtual reality platform. This study is first to show that the percentile-matching technique enables more accurate approximation of the target set using the biometric user authentication as an end task. Its mean absolute error of estimating the target set performance level is better for all three metrics considered (compared to the baseline) - a) 0.26 (versus 0.84) of d-prime; b) 12.08% (versus 31.10%) of False Rejection Rate at 1-in-10000 False Acceptance Rate; c) 3.08% (versus 4.69%) of Equal Error Rate. This paper also expands eye movement literature by suggesting an application-agnostic realism assessment workflow based on the 1 Nearest-Neighbor classifier. Our model improves the related median classification accuracy metric by 35.7% versus the benchmark model towards the ideal 50% value.
Data augmentation is commonly utilized to increase the size and diversity of training sets for deep learning tasks. In this study, we propose a novel application of an existing image generation approach in the domain of realistic eye images that leverages data collected from 40 subjects. This hybrid method combines the benefits of precise control over the image content provided by 3D rendering, while introducing the previously lacking photorealism and diversity into synthetic images through neural style transfer. We prove its general efficacy as a data augmentation tool for appearance-based gaze estimation when generated data are mixed with a sparse train set of real images. It improved the results for 39 out of 40 subjects, with an mean and a maximum decrease in gaze estimation error, achieving similar metrics for train and held-out subjects. We release our data repository of eye images with gaze labels used in this work for public access.