Pose estimation of anime/manga characters

Pramook Khungurn, Derek Chou
2016 Proceedings of the 1st International Workshop on coMics ANalysis, Processing and Understanding - MANPU '16  
2D articulated pose estimation is the task of locating the body joint positions of a human figure in an image. A pose estimator that works on anime/manga images could be an important component of an automatic system to create 3D animation from existing manga or anime, which could significantly the lower cost of media production. To create an accurate pose estimator, however, a sizable and high-quality dataset is needed, and such a dataset can be expensive to create. To alleviate data scarcity,
more » ... e propose using a database of 3D character models and poses to generate synthetic training data for 2D pose estimators of anime/manga characters based on convolutional neural networks (CNN). We demonstrate that a high-performing estimator can be obtained by pretraining a network on a large synthetic dataset and then fine-tuning it on a small dataset of drawings. We also show that the approach yields a pose estimator competitive with many previous works when applied to a photograph-based dataset, establishing our synthetic data's usefulness beyond the intended domain.
doi:10.1145/3011549.3011552 dblp:conf/icpr/KhungurnC16 fatcat:edhcuyjszrf4lezchum5whu4ma