Word-level Stroke Trajectory Recovery for Handwriting with Gaussian Dynamic Time Warping

Anonymous

Abstract: Handwriting trajectory recovery has recently gained more attention for practical applications such as personalized messages. It is a sequence learning problem from image to handwriting stroke sequence where Dynamic Time Warping (DTW) is a preferred loss function. However, aligning two varying length sequences in DTW loss accumulates the differences of predicted and ground truth strokes for the entire line-level text. As a result, averaging over long sequences in DTW loss, it cannot distinguish between a small number of perceptually significant errors and a large number of visually insignificant errors. To address this issue, we propose two new strategies. First, we propose applying DTW to words instead of line-level text so that the DTW loss for all the words in the line-level text is not averaged out. Moreover, for aligning the predicted and ground-truth sequences for each word, we propose to weight the cost matrix with a Gaussian function so that the far-off predicted strokes from ground truth are penalized heavily. This strategy for word-level stroke trajectory learning improves quantitative and qualitative results.

Paper Type: short