Keynote Speakers

Prof. Xiao Wu

Keynote Speaker 1: Prof. Xiao Wu

Southwest Jiaotong University

Biography

Xiao Wu is a Professor of Southwest Jiaotong University, Chengdu, China. He received the Ph.D. degree in Computer Science from City University of Hong Kong, Hong Kong. He was with the School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, USA, and at the School of Information and Computer Science, the University of California, Irvine, CA, USA, as a Visiting Scholar from 2006 to 2007 and from 2015 to 2016, respectively. His research interests include artificial intelligence, computer vision, and intelligent transportation systems. He has authored or co-authored more than 150 research papers in well-respected journals, such as TIP, TMM, and TMI, and prestigious proceedings like CVPR, ICCV, and ACM MM. He received Second Prize of Natural Science Award of Sichuan Province, China, in 2025, the Second Prize of Natural Science Award of the Ministry of Education, China, in 2016, the Second Prize of Science and Technology Progress Award of Henan Province, China, in 2017, the Best Paper Award of International Conference on Multimedia Modeling, in 2021, and the Best Paper Award of ACM International Conference on Multimedia Asia in 2024.

Title: Human-Centered Interaction Detection and Applications

Abstract: Understanding human-centered interactions is pivotal for developing intelligent systems capable of perceiving, interpreting, and responding to human activities in complex environments. This presentation will provide a comprehensive overview of advanced research in human-object interaction (HOI) detection, extending the scope to encompass critical interaction paradigms such as human-object-human (HOH), human-object-object (HOO) and contactless human-object interactions. The core focus will be on elucidating the methodologies, challenges, and practical applications of these interaction detection technologies within industrial settings.


Prof. Junyu Gao

Keynote Speaker 2: Prof. Junyu Gao

Institute of Automation, Chinese Academy of Sciences (CAS)

Biography

Junyu Gao is an Associate Professor at the Institute of Automation, Chinese Academy of Sciences (CAS), and a recipient of a national-level young talent award. His research interests include embodied intelligence, video understanding, and multimedia computing. He has published over 60 papers in flagship conferences and leading journals in artificial intelligence, including more than 20 first-author papers, with over 4,700 citations on Google Scholar. He has received several prestigious honors, including the CAS President's Special Award, the CAS Outstanding Doctoral Dissertation Award, and the Baidu Scholarship. He was also selected for the Young Elite Scientists Sponsorship Program by the China Association for Science and Technology (CAST). As a principal investigator or project leader, he has led numerous research projects funded by the National Natural Science Foundation of China (NSFC), including the Young Scientists Fund (Category B), the General Program, and key research tasks under major NSFC projects. His research achievements have been recognized with several prestigious awards, including the First Prize of the 2025 Wu Wenjun Artificial Intelligence Science and Technology Award for Technological Invention, the Second Prize of the 2024 Jiangsu Provincial Science and Technology Award, and the Best Student Paper Award at ACM Multimedia 2025.

Title: Building Brain-Inspired World Models for Embodied Navigation and Tracking

Abstract: Embodied intelligence has advanced rapidly, driven by recent developments in multimodal learning, robotics, and cognitive science, demonstrating tremendous potential for enabling human-centric interaction and task execution. Embodied mobility is a fundamental component of embodied intelligence, encompassing two primary tasks: navigation and tracking. The former aims to guide an embodied agent to a designated location according to user instructions, while the latter enables the agent to continuously follow a target of interest. In this talk, we first introduce a novel learning and evolution paradigm for embodied navigation agents. Specifically, we develop an embodied context-enhanced multimodal large language model (MLLM) to emulate the language processing and logical reasoning capabilities of the brain's left hemisphere, enabling effective instruction understanding and visual scene interpretation. Meanwhile, we construct a perceptual context-guided world model based on a recurrent state-space model to mimic the spatial perception and holistic reasoning functions of the right hemisphere, thereby capturing environmental dynamics and predicting future states. Extensive experimental results demonstrate that the proposed paradigm significantly improves the performance of embodied agents across a wide range of tasks in zero-shot scenarios. Building upon this, we further propose a Bipartite World Modeling Framework that decomposes the latent state into two complementary components: an egocentric state that models global scene evolution and an allocentric state that maintains temporally coherent target representations. These two states are jointly inferred and evolve through differentiated transition mechanisms within a unified world model. We evaluate the proposed framework on embodied target tracking benchmarks, a representative setting characterized by heterogeneous dynamics, and further validate its effectiveness through real-world robotic deployment. Extensive experiments on three benchmarks demonstrate that our approach achieves state-of-the-art performance.


Prof. Yicong Li

Keynote Speaker 3: Prof. Yicong Li

University of Science and Technology of China

Biography

Yicong Li is a Specially Appointed Professor at the University of Science and Technology of China (USTC) and a recipient of the National Excellent Young Scientists Fund (Overseas). He received his Ph.D. from the National University of Singapore under Prof. Tat-Seng Chua and his M.S. from Columbia University. His research focuses on efficient multimodal learning and generation, large language models, and multi-agent systems. His work has been recognized as a CVPR 2022 Best Paper Finalist.

Title: Generate More with Less: Efficient Generation from Images to Video to 4D

Abstract: Generative models have moved from images to video and on to dynamic 4D scenes, and each added dimension makes generation more expensive, in both compute and data. This talk argues that much of this cost is not intrinsic to the task, and addresses it from two angles. The first is mismatch: when a model treats its task as something it is not, the difference is paid for with larger models, more sampling steps, or redundant representations. Fixing the mismatch, rather than scaling up, recovers that compute. The second is data: the paired data a new task needs often does not exist. By recasting tasks as editing, we can reuse data and models we already have, such as ordinary videos and pretrained generators, instead of collecting new data. I will show how both ideas apply across image, video, and 4D generation, and close with interactive generation, where both matter even more under real-time constraints.