Jan 26, 2024 Friday
Abstract:
As a basic 3D representation form, point cloud is active in various tasks such as autonomous driving, robot perception, biomolecular structure prediction and design. Although the 3D point cloud analysis has achieved good development in recent years. Considering the large scene point cloud itself is often a huge amount of data, and has the characteristics of disorder, no texture and sparseness, the algorithm of large scene point cloud analysis is recently has been the focus of recent research. Starting from the collection of point clouds, this report first proposes a point cloud down-sampling and recovery algorithm based on reversible networks, which greatly improves the storage and communication efficiency of point cloud data in large scenes. After we can effectively obtain point cloud data, we have studied classic tasks such as point cloud shape classification, point cloud 3D detection and tracking, and large scene 3D semantic segmentation, from single modality to multimodal fusion and distillation, Our algorithm has achieved excellent results in many public international competitions, such as the first place in SemanticKitti semantic segmentation, the first place in CVPR2023 HOI4D segmentation, the second place in ICCV21 Urban 3D competition, the third place in NuScenes semantic segmentation, etc. In the end, we extended the analysis algorithm of large-scale scene point clouds to downstream applications, such as visual reasoning of 3D scenes, description generation of 3D scenes, face generation of AI anchors, prediction of binding of small protein molecules and binding sites, etc.
Time:
Jan 26, 2024 Friday
11:00-11:50
Location:
Rm W1-101, GZ Campus
Online Zoom
Join Zoom at: https://hkust-gz-edu-cn.zoom.us/j/4236852791 OR 423 685 2791
Speaker Bio:
Dr. ZHEN LI 李镇
Assistant Professor, SSE & FNII, The Chinese University of Hong Kong (Shenzhen)
Dr. Li Zhen is currently an Assistant Professor and Principal Young Scholar at the School of Science and Engineering (SSE)/Future Network Intelligent Institute (FNII), the Chinese University of Hong Kong, Shenzhen (CUHKSZ). Dr. Zhen Li received his PhD in Computer Science from the University of Hong Kong (2014-2018), and he also served as a visiting scholar at the University of Chicago in 2018. Dr. Li Zhen won the 7th Youth Lifting Talent of China Association for Science and Technology in 2021, the global champion of CASP12 contact map prediction, the first place in CVPR2023 HOI4D segmentation competition, the first place in SemanticKITTI competition in 2022, the second place in Urban3D competition 2021, and the Urban3D competition 2022 third place. Dr. Li Zhen has also won scientific research projects from the national, provincial and municipal levels and the industry. Dr. Li Zhen led the Deep Bit Lab (https://mypage.cuhk.edu.cn/academics/lizhen/) in CUHKSZ. His main research direction is 3D visual analysis and application (including but not limited to point cloud analysis, multi-modal joint analysis), deep learning and other basic theoretical algorithm research, and is committed to Published more than 60 papers in well-known international journals and conferences in this direction, including top journals Cell Systems, Nature Communications, Bioinformatics, in applying 2D/3D artificial intelligence algorithms to protein/RNA structure prediction, autonomous driving, industrial vision and other scenarios , TMI, TNNLS, etc., and top conferences CVPR, ICCV, ECCV, NeurIPS, AAAI, IJCAI, MM, MICCAI, Recomb, etc. Dr. Li Zhen serves as the deputy editor of IEEE Transactions on Mobile Computing (TMC), IROS and the reviewer of many top journals and conferences.