|14:45-15:30||Psychology-Inspired Visual Emotion Analysis||杨景媛|
|15:30-16:15||Wide-Area Crowd Counting via Deep Learning based Multi-View Fusion||张琦|
|17:00-17:45||Architectural design and geometry optimization||熊卫丹|
Psychology-Inspired Visual Emotion Analysis
Visual Emotion Analysis (VEA) aims at finding out how people feel emotionally towards different visual stimuli, which has attracted great attention recently with the prevalence of sharing images on social networks. Although researchers have been engaged in VEA for more than two decades, there remain several challenges to be solved, including abstractness, ambiguity, and subjectivity. Rather than simply implementing a universal effective network to VEA, we argue that it is more reasonable to design emotion-specific networks with prior knowledge from related fields. Inspired by psychological studies, we propose a series of methods to solve the three core challenges in VEA, aiming at simulating the human emotion evocation process with deep neural networks. 1) Abstractness: where do emotions come from? To be specific, we first propose the stimuli-aware network to predict emotions from visual stimuli, including color, object, and face. Moreover, we construct a scene-object interrelated network to mine visual emotions from objects and scenes. 2) Ambiguity: how do emotions interact with each other? We propose a circular-structured representation by exploiting the intrinsic relationship between different emotions. 3) Subjectivity: why different people have different emotions? We propose a subjectivity appraise-and-match network with affective memory to depict the diversity in crowd emotion evocation process. In this presentation, we will introduce the above psychology-inspired VEA solutions in detail.
Jingyuan Yang received the B.Eng. degree in electronic and information engineering from Xidian University, Xian, China, in 2017, where she is currently pursuing the Ph.D. degree with the School of Electronic Engineering. Her current research interests include visual emotion analysis, computer vision and deep learning.
Wide-Area Crowd Counting via Deep Learning based Multi-View Fusion
Deep learning-based crowd counting in single-view images has achieved outstanding performance on existing counting datasets. However, single-view counting is not applicable to large and wide scenes (e.g., public parks, long subway platforms, or event space) because a single camera cannot capture the whole scene in adequate detail for counting, e.g., when the scene is too large to fit into the field-of-view of the camera, too long so that the resolution is too low on faraway crowds, or when there are too many large objects that occlude large portions of the crowd. Therefore, to solve the wide-area counting task requires multiple cameras with overlapping fields-of-view. Traditional multi-view counting relies on foreground extraction techniques and hand-crafted features, which limit the multi-view counting performance. To address these problems, we explore deep learning-based multi-view fusion methods for better wide-area crowd counting performance.
张琦，博士毕业于香港城市大学电脑科学系，主要研究方向是基于多相机融合的人群计数和视频监控问题。本科毕业于华中科技大学光电信息学院（光电信息工程专业），后硕士保送至中国科学院大学（信号与信息处理专业），培养单位为中科院西安光机所。近三年已发表一作CVPR 2篇，AAAI 1篇，授权美国/中国发明专利各一项，参与香港研究资助局资助科研项目3项。担任CVPR 2022/2021/2020，AAAI 2022/2021, ICCV 2021,ICPR 2020, WACV 2022/2021/2020等,和IEEE TIP, TCSVT, TMM 等多个主流CV会议和期刊审稿人。
随着3D娱乐应用的普及和网络服务的快速发展，研究视觉显著性以区分3D场景中的显著区域的兴趣急剧增长。本次讲座旨在讲述基于先进的深度神经网络的立体 视频和光场等3D场景的视觉显著性检测研究。研究内容主要由3个部分组成：1）通过探索立体视频中时间、空间和深度线索的关系，开发了一种基于组件交互的立体视频视觉显著性检测模型；2) 联合边缘检测、深度推理和显著性对象检测进行多任务协同学习，开发了一种基于多任务协同网络的光场显著性对象检测模型；3）通过使用图卷积神经网络探索光场的几何特性，开发了一种基于几何辅助的光场显著性对象检测模型。这三部分分别对3D场景的视觉显著性检测进行了深入研究，为未来3D场景视觉显著性检测的综合注意力模型提供了有益的见解。
张秋丹，2021年获得香港城市大学计算机科学系博士学位，2018年获得深圳大学硕士学位，2015年本科毕业于深圳大学。主要研究方向是3D计算机视觉，多媒体图像和视频内容处理分析，基于深度学习的3D视觉感知等。研究成果均发布于IEEE TIP、TCSVT和IEEE CVPR等国际期刊及会议。
Architectural design and geometry optimization
Architecture has a long history, and humans have tried to build complex architectural forms satisfying functional requirements as well as aesthetic ones. The design and construction cycle of such complex structures is non-trivial. Many sub-tasks are often completed by tedious manual work. Furthermore, the design obtained by such efforts may not be optimal from the point of view of constructions costs, risks, or even functionality. Various modeling, analysis, and optimization techniques have been proposed to improve the level of design automation and to accelerate the design and construction processes. In this talk, we will introduce an interactive architectural design interface, along with optimization algorithms to focus on sub-problems in the lifecycle.
Dr. Weidan Xiong received her PhD degree in the Department of Electronic and Computer Engineering at The Hong Kong University of Science and Technology in 2019. She is currently a research fellow in Nanyang Technological University. Her research interests include computer graphics, CAD and machine learning.