Origami is a recreational activity that involves folding a paper to create shapes imitating various objects. To facilitate the sharing and instruction of origami creations, folding diagrams and instructional videos are widely employed to illustrate the folding procedure. However, since creating folding diagrams requires significant time and effort, we are attempting to automatically generate them from instructional videos. In this paper, we first decompose the generation process into a series of sub-problems. Then, we focus on one of the sub-problems, the crease estimation, which is the task to estimate a folding line (i.e., crease) from a pair of images taken before and after a folding operation. This paper is the first attempt to address this task using a learning-based methodology. Evaluation on a hand-crafted dataset showed that a model using a five-layer CNN as the backbone achieved the best performance, with a distance displacement of approximately 6% of the image’s side length and an angle displacement of approximately \(14^\circ \) . These results suggest the potential feasibility of constructing a learning-based crease estimator.