Evaluation of Gaze-to-Object Mapping Algorithms for Use in "Real-World" Translatable Neuropsychological Paradigms

Eye-tracking technology is commonly used for identifying objects of visual attention. However, applying this technology to virtual reality (VR) applications is challenging. This report analyzes the performance of two gaze-to-object mapping (GTOM) algorithms applied to eye-gaze data acquired during a "real-world" VR cue-reactivity paradigm.

Two groups of participants completed a VR paradigm using an HTC Vive Pro Eye. The gazed objects were determined by the reported gaze rays and one of two GTOM algorithms - naïve ray-casting (n=18) or a combination of ray-casting and Tobii's G2OM algorithm (n=18). Percent gaze duration was calculated from 1-second intervals before each object interaction to estimate gaze accuracy. The object volume of maximal divergence between algorithms was determined by maximizing the difference in Hedge's G effect sizes between small and large percent gaze duration distributions. Differences in percent gaze duration based on algorithm and target object size were tested with a mixed ANOVA.

The maximum Hedge's G effect sizes differentiating large and small target objects was observed at an 800cm³ threshold. The combination algorithm performed better than the naïve ray-casting algorithm (p=.003, η_p²=.23), and large objects (>800cm³) were associated with a higher gaze duration percentage than small objects (≤800cm³; p<.001, η_p²=.76). No significant interaction between algorithm and size was observed.

Results demonstrated that Tobii's G2OM method outperformed naïve ray-casting in this "real-world" paradigm. As both algorithms show a clear decrease in performance for detecting objects with volumes <800cm³, we recommend using gaze-interactable objects >800cm³ for future HTC Vive Pro Eye applications.