(英) |
Research on 3D pose estimation from photographs of real human using machine learning have already been proposed. However, because these methods focus on photographs of real human, when it apply to illustrations of anime character, it will be low estimation accuracy. To this end, we investigate the difference in head-body ratio of each character, which is one of the characteristics of illustrations of anime character, and propose an improved method for illustrations of anime character. Through expanding dataset by changing not only the poses, but also the head-body ratios, our approach can be applied to different head-body rate. The experiments result shows that our approach can reach a higher estimation accuracy in rendered images of 3D models. |