CASE: Context Aware Screen-Based Estimation of Gaze

High quality human eye gaze estimation in the 0.43° to 1.5° range is achievable, but only by approaches with high-end specialized hardware such as infrared sensor bars or sensor-enabled glasses. For many low-cost applications this barrier to entry deters or rules out eye gaze entirely as a method of inferring human intention. More accessible deep learning webcam-based approaches have been proposed, but trail behind significantly in accuracy with values greater than 3° common. Here, we propose CASE, a deep learning gaze estimation pipeline that supplements existing webcam-based estimation methods with screenshots of current and historical screen content. Our unique demand for additional data limits our current ability to benchmark results against most existing gaze datasets. As we work on collecting a more robust dataset that may fill this gap, we express our optimism in this approach's ability to enable accessible higher-quality gaze estimation and detail the current state of our methods. Code will be available at github.com/SpamMusubi153/CASE.