Abstract
Existing correspondence datasets for two-dimensional (2D) cartoon suffer fromsimple frame composition and monotonic movements, making them insufficient tosimulate real animations. In this work, we present a new 2D animation visualcorrespondence dataset, AnimeRun, by converting open source three-dimensional(3D) movies to full scenes in 2D style, including simultaneous movingbackground and interactions of multiple subjects. Our analyses show that theproposed dataset not only resembles real anime more in image composition, butalso possesses richer and more complex motion patterns compared to existingdatasets. With this dataset, we establish a comprehensive benchmark byevaluating several existing optical flow and segment matching methods, andanalyze shortcomings of these methods on animation data. Data, code and othersupplementary materials are available athttps://lisiyao21.github.io/projects/AnimeRun.