报告时间：2025年7月21日（星期一）14:55
报告地点：翡翠湖校区科教楼A1104室
报 告 人：Dr. Hongjie He
工作单位：University of Waterloo
举办单位：计算机与信息学院（人工智能学院）
报告简介：
High-definition building mapping is essential for geospatial applications, but deep learning methods typically require extensive annotated datasets. This seminar presents Box2Boundary, a novel weakly supervised approach that integrates box-level supervision with the Segment Anything Model (SAM) to achieve precise building segmentation while dramatically reducing annotation costs compared to conventional fully supervised methods. The seminar will begin with an overview of weakly supervised learning in remote sensing. We will then present the Box2Boundary methodology and discuss its evaluation on three publicly available building datasets. The seminar will conclude with a discussion of future research directions, particularly the integration of foundation models like SAM into scalable geospatial mapping pipelines and the broader potential for cost-efficient large-scale urban analysis.
报告人简介：
Dr. Hongjie He received his Ph.D. in Geography from the University of Waterloo in 2023 and is currently a postdoctoral fellow working with Professor Jonathan Li. His research focuses on AI-based algorithms for information extraction from earth observation images. He has published papers in the International Journal of Applied Earth Observation and Geoinformation, IEEE Transaction on Geoscience and Remote Sensing, Canadian Journal of Remote Sensing and Geomatica, flagship conferences including IGARSS and ISPRS. He is a member of the IEEE Geoscience and Remote Sensing Society (GRSS), the Canadian Institute of Geomatics (CIG), and the Canadian Remote Sensing Society (CRSS). He also serves as a Lead Guest Editor for the Remote Sensing journal.