Eye behaviour has gained much interest in the VR research community as an interaction input and support for collaboration. Researchers implemented gaze inference models when eye-tracking is missing by using head behavior and saliency. However, these solutions are resource-demanding and thus unfit for untethered devices, and their angle accuracy is around 7°, which can be a problem in high-density informative areas. To address this issue, we propose a lightweight deep learning model that generates the probability density function of the gaze as a percentile contour. This solution allows us to introduce a visual attention representation based on a region rather than a point and manage a trade-off between the ambiguity of a region and the error of a point. We tested our model in untethered devices with real-time performances; we evaluated its accuracy which outperforms our identified baselines (average fixation map and head direction).