This article is automatically translated.
In this article, I have summarized some of the research presented at SIGGRAPH Asia 2018. In a previous article, we introduced the content exhibited by Dwango Media Village at SIGGRAPH Asia 2018 and provided an overview of the conference.
Here are some of the papers presented at SIGGRAPH Asia 2018 that I found particularly interesting. The research focused on improving accuracy in problems like 3D reconstruction, face deformation, and depth estimation by adding new insights to existing methods. These ideas can be applied to other tasks, making the research quite fascinating. Given that the conference was held in Tokyo, many poster sessions featured uniquely Japanese problems, including a study on rendering and animating Japanese-style cel animations in 3D, which I found particularly notable.
This paper discusses semi-automatic coloring tasks. I found it particularly interesting as it cited a paper I had submitted to SIGGRAPH Asia last year. Existing illustration coloring methods often suffer from color bleeding issues. This problem was addressed by dividing the task into two steps: a rough coloring step (draft) and a refinement step (improvement). The research reported that this method achieves cleaner coloring compared to existing techniques.
These two studies were part of the Image Processing session, where most of the research was based on Cycle GAN. This highlights the high level of interest and ongoing research in unpaired learning methods and style transfer.
From the Character Animation session at the SIGGRAPH Asia 2018 CG conference held in December 2018, I have summarized two studies that use reinforcement learning to automatically generate 3D character motions. Creating character motions manually requires a significant amount of effort, but using reinforcement learning, an agent can automatically learn how to move through simulations. However, it remains challenging to learn complex and long motions like backflips and dressed movements in environments that simulate physical phenomena like gravity. One paper addresses this by using dynamic initial state sampling, while the other approaches it by breaking down the motion into shorter segments.
At Dwango Media Village, we not only participated in SIGGRAPH Asia 2018 but also share papers and previously presented materials on platforms like Niconare and GitHub issues. Feel free to explore and make use of these resources.
Chie Furusawa
Keisuke Ogaki
Kazuyuki Hiroshiba
Kazuma Sasaki
