第8期图像图形学科前沿讲习班11月17日杭州举办

刘洋

微软亚洲研究院 主管研究员

报告题目：从想象到现实：从草图出发构建三维自由曲面

摘要：设计者通常在纸面上勾勒二维线条或在三维空间绘制稀疏的三维线框以快速描绘想象中的三维形体，但如何将想象中的三维形体的曲面构造出来一直是个未解决的问题。此次报告将展示通过绘制稀疏二维或三维草图来创造三维自由曲面的技术。该技术将表达物体形状的草图形式和期望曲面的曲率流向及特征方向相关，由此为基础设计了数值上高效鲁棒的计算方法以构建符合用户设计意图的三维曲面。

Charlie C.L. Wang

香港中文大学 教授

报告题目：Geometric Computing for Multi-Axis Additive Manufacturing

摘要：In this talk, I will present our recent development of using multi-axis motion to conduct material accumulation along dynamically varied directions. Our development results in two approaches that mainly focus on how to avoid the additional supporting structures in a framework of volume-to-surface and then surface-to-curve decomposition. I will discuss a few future extensions of this framework so that models can be printed faster and in a more accurate way.

Tien-Tsin Wong

香港中文大学 教授

报告题目：Learning for Graphics When Training Data is Scarce

摘要：Deep learning has been demonstrated to be an effective tool for solving many problems that are ambiguous in nature. It outperforms many tailormade solutions and offers stable results, in a real-time speed. It seems to be an ultimate solution for many hard problems. However, its major drawback is its high dependency on training data, because it transforms the problem from “method” to “data.” In many cases, training data is scarce or hard/impractical to obtain.

In this talk, I will talk about a few learning-for-graphics projects we have done in the past few years at CUHK. In some of these projects, we face the problem of lacking groundtruth training data. We have worked on computational manga for many years. A key problem we want to solve since the beginning is to remove the screentones that manga artists typically used to enrich the manga. Removing the screentones can significantly ease the digital migration from paper to digital platform. However, the definition of screentones by itself is ambiguous and cannot be easily defined by mathematical equations. More importantly, we do not have the unscreened manga as groundtruth, as it is prohibitively expensive to manually trace the structural lines from legacy manga. We shall describe how we overcome the problem of data scarcity. I will discuss how to extend our strategy to other applications such as sketch colorization, and also unsupervised learning framework for invertible grayscale application, in this talk.

刘世霞

清华大学软件学院 长聘副教授

报告题目：交互式机器学习模型分析

摘要：可解释的机器学习旨在使机器学习模型的决策过程对研究人员和从业人员更加透明，从而实现人机的有效沟通和协作。本报告将介绍我们提出的机器学习模型可视分析框架。该框架跳出传统可视分析“先分析再可视化”的单一方向分析机制，将机器学习方法和交互可视化方法有机地结合在一起，从而更好地帮助用户理解复杂模型及其输出结果，分析、诊断并不断完善机器学习模型。为用户选择、利用及改进机器学习模型提供技术依据。最后，结合具体的应用实例，如集成学习模型和深度学习模型分析等，介绍我们基于该框架研制开发的可视分析技术。

廖菁

香港城市大学计算机科学系 助理教授

报告题目：基于深度学习的艺术风格化研究

摘要：运用人工智能（AI）进行艺术创作一直是极具挑战的研究领域。近年来深度学习的技术极大地加速了这些领域的发展，如微软小冰作诗，谷歌大脑音乐创作，Prisma艺术风格化滤镜等。本次讲座我们将讨论如何将传统的视觉图形学研究和深度学习技术有机地融合，从而有效地进行图片和视频的艺术风格化，使得普通的生活照片变身为大师级绘画或摄影作品。

讲座嘉宾将介绍其近两年基于深度神经网络学习风格迁移（style transfer）的一系列科研成果。首先介绍如何设计深度神经网络学习语义信息和艺术风格，然后将其艺术风格根据语义应用到新的照片，视频，VR/AR中; 其次介绍一些相关的扩展应用包括照片颜色转换（比如白天变黑夜、夏季变冬季），人像化妆迁移，黑白照片上色等。

徐枫

清华大学软件学院 特别研究员

报告题目：从人到场景的动态三维重建

摘要：在计算机视觉中，实时准确的获取人和场景的各种运动信息，将成为未来自然人机交互的技术基础；与此同时，在计算机图形学中，高精度的重建运动，对于电影、游戏中的特效生成与制作具有重要意义。本报告首先围绕人体动态重建问题介绍若干最新的研究工作，包括利用不同的硬件设备，针对高精度、便捷性和实时性三大目标，对人体的不同部分（包括面部，肢体，目光）进行运动重建。之后，针对360度场景的动态重建技术，介绍课题组的最新研究成果。