Mixture-of-Experts Improves Digital Pen Handwriting Reconstruction.

Digital pens equipped with sensors are becoming crucial for human-computer interaction, allowing users to write on any surface while preserving a digital trace. A key challenge is accurately reconstructing the handwriting trajectory, especially distinguishing between pen-on-surface (touching) and pen-in-air (hovering) movements. This research proposes a new approach utilizing a Mixture-Of-Experts (MOE) model to address this. One expert model is specifically trained for the touching phase of the pen, while another is dedicated to the hovering trajectory. This specialized learning allows for finer reconstruction of touching traces and more precise positioning for subsequent strokes. The study demonstrates significant improvements over existing methods and introduces a new public benchmark dataset. This dataset will facilitate future research and comparisons in the field of handwriting reconstruction, particularly for applications like aiding writing education.