Two-dimensional(2D)character animation is one of the most important visual elements on which users’interest is focused in the game field.However,2D character animation works in the game field are mostly performed man...Two-dimensional(2D)character animation is one of the most important visual elements on which users’interest is focused in the game field.However,2D character animation works in the game field are mostly performed manually in two dimensions,thus generating high production costs.This study proposes a generative adversarial network based production tool that can easily and quickly generate the sprite images of 2D characters.First,we proposed a methodology to create a synthetic dataset for training using images from the real world in the game resource production field where machine learning datasets are insufficient.In addition,we have enabled effective sprite generation while minimizing user input in the process of using the tool.To this end,we proposed a mixed input method with a small number of segmentations and skeletal bone paintings.The proposed image-to-image translation network effectively generated sprite images from the user input images using the skeletal loss.We conducted an experiment regarding the number of images required and showed that 2D sprite resources can be generated even with a small number of segmentation inputs and one skeletal bone drawing.展开更多
基金This work was supported by the National Research Foundation of Korea(NRF)grant funded by the Korean Government(MSIT)(Nos.NRF-2019R1A2C1002525,NRF-2020R1G1A1100125)a research grant from Seoul Women’s University(2020-0181).
文摘Two-dimensional(2D)character animation is one of the most important visual elements on which users’interest is focused in the game field.However,2D character animation works in the game field are mostly performed manually in two dimensions,thus generating high production costs.This study proposes a generative adversarial network based production tool that can easily and quickly generate the sprite images of 2D characters.First,we proposed a methodology to create a synthetic dataset for training using images from the real world in the game resource production field where machine learning datasets are insufficient.In addition,we have enabled effective sprite generation while minimizing user input in the process of using the tool.To this end,we proposed a mixed input method with a small number of segmentations and skeletal bone paintings.The proposed image-to-image translation network effectively generated sprite images from the user input images using the skeletal loss.We conducted an experiment regarding the number of images required and showed that 2D sprite resources can be generated even with a small number of segmentation inputs and one skeletal bone drawing.