Make Gaze Data Great Again for Panoramic Scenes: A Progressive Iterative Self-Paced Learning

When analyzing panoramic images, the trajectory of a user's head or eye movements, known as gaze data, is crucial to understanding their exploration of the entire scene. However, preprocessing gaze data in panoramic scenes presents unique challenges compared to traditional 2D images. Conventional methods for regular 2D images heavily rely on hand-crafted low-level features, such as velocity and dispersion, which are ill-suited for the rich information contained in panoramic images. Panoramic scenes often require frequent transitions between fixations and saccades, unlike 2D scenes, making it difficult for existing methods to accurately differentiate meaningful fixations from meaningless saccades based solely on low-level features. To address this issue, we propose a novel approach for gaze data preprocessing in panoramic scenes, leveraging progressively iterative self-paced learning to dynamically identify meaningful fixations. The key innovation of our approach is to focus on more informative areas (i.e., fixations) and filter out meaningless areas (i.e., saccades), resulting in fixation sequences that align better with human viewing habits in panoramic environments. Additionally, since this is the first attempt at gaze data preprocessing in panoramic scenes, we introduce a set of evaluation metrics to assess the effectiveness of our approach and compare it with other methods. The results demonstrate that our method achieves superior performance compared to gaze data preprocessing approaches in this task.