VALSE Webinar 23-31期 总第331期 图像合成前沿进展

报告嘉宾：张赫 (Adobe Research)

报告题目：Towards Next-Generation Image Composition (1)

报告嘉宾：张健明 (Adobe Research)

报告题目：Towards Next-Generation Image Composition (2)

报告嘉宾：牛力 (上海交通大学)

报告题目：Data-centric Image Composition

报告嘉宾：张赫 (Adobe Research)

报告时间：2023年11月29日 (星期三)晚上20:00 (北京时间)

报告题目：Towards Next-Generation Image Composition (1)

报告人简介：

He Zhang is a senior research scientist at Adobe Research. where his main focus at Adobe is to design the workflow for image composition including high-quality object segmentation, image harmonization, object/ portrait relighting, object/ portrait shadow generation and inverse rendering for image editing using diffusion model.  He was a computer vision engineer working at Apple Vision Pro. He did his PhD advised by Vishal M. Patel.

报告摘要：

In the first part of the talk, I will discuss about recent progress in the area of image composition, where Image composition refers to the process of combining multiple visual elements to create a new and cohesive composition. To achieve such high-quality cohesive composition, we’ll mainly discuss about the recent research progress on the following topics including 1. High-quality segmentation, 2. Image Harmonization, 3. Relighting and their combinations with new GenAI (diffusion model) technology.

报告嘉宾：张健明 (Adobe Research)

报告时间：2023年11月29日 (星期三)晚上20:30 (北京时间)

报告题目：Towards Next-Generation Image Composition (2)

报告人简介：

Dr. Jianming Zhang is a computer vision researcher at Adobe Research. He obtained PhD in computer science from Boston University in 2016. His research interests include image analysis, image synthesis and photo editing. He has published over 90 papers in peer-reviewed conferences and journals. He served as area chair for CVPR 2022.

报告摘要：

In the second part of the talk, I will present our latest work in improving object compositing workflows. We are exploring two main approaches: 1) synthesizing and harmonizing light effects based on geometry, and 2) using generative inpainting for implicit object compositing. Specifically, we will cover our research work on PixelHt Lab, a light effect synthesis system based on pixel height, and ObjectStitch, a diffusion-based method for object compositing. We will discuss the advantages and disadvantages of each method and consider potential ways to integrate both approaches.

报告嘉宾：牛力 (上海交通大学)

报告时间：2023年11月29日 (星期三)晚上21:00 (北京时间)

报告题目：Data-centric Image Composition

报告人简介：

牛力现为上海交通大学电子信息与电气工程学院计算机科学与工程系长聘教轨副教授，本科毕业于中国科学技术大学，博士毕业于新加坡南洋理工大学。主要研究领域是计算机视觉，以第一作者或通讯作者身份在计算机视觉和人工智能顶级会议 (CVPR, ICCV, ECCV, NeurIPS等)和知名期刊 (IJCV, TIP 等)上发表论文70余篇，近三年的工作主要集中于图像编辑，公布了image composition领域的第一代数据集，其中包括图像和谐化领域最常用的数据集iHarmony4，推出了image composition领域的首个集成工具箱libcom。

报告摘要：

Image composition aims to combine a foreground image and a background image to generate a realistic composite image, which has a wide range of applications in virtual reality, entertainment, artistic creation, automatic advertising, and so on. The issues of unrealistic composite images can be summarized as the inconsistency between foreground and background, including appearance inconsistency, geometric inconsistency, and semantic inconsistency. Some works focus on only one issue, while some other works attempt to address multiple issues parallelly or sequentially, leading to different tasks named image harmonization, shadow generation, object placement, foreground object search, generative image composition. This talk will briefly introduce different tasks and our representative works in the field of image composition.

主持人：郭宗辉 (中国科学院计算技术研究所)

主持人简介：

郭宗辉，中国科学院计算技术研究所智能信息处理重点实验室博士后，主要从事图像合成与伪造检测方向研究，以第一作者在IEEE TPAMI、CVPR、ICCV等期刊和会议上发表学术论文6篇，申请发明专利2项，获中国海洋大学研究生卓越奖学金，主持国家自然科学基金面上项目1项，任VALSE EACC，VALSE年度会议注册主席 (2019至今)。

特别鸣谢本次Webinar主要组织者：

主办AC：郭宗辉 (中国科学院计算技术研究所)