Geoinformatics Team Seminar (Talk by Jian Song, The University of Tokyo).

Title: From Synthetic to Real: Low-Cost Pathways for Semantic 3D Understanding in Remote Sensing

Abstract:
Achieving robust semantic 3D understanding in remote sensing with deep learning is hindered by limited data, costly annotations, and the domain gap introduced by geographic biases. In this talk, I will present a novel pipeline that automates the generation of vast, lifelike synthetic cityscapes—supplying reliable yet low-cost training data for tasks such as land cover mapping and height estimation. To mitigate the domain shift and further enhance synthetic data usability, we propose the first syn2real multi-task unsupervised domain adaptation framework in remote sensing, significantly improving model performance and stability. Finally, I will illustrate how sparse, high-precision measurements from real-world can refine these deep learning models' prediction, boosting both accuracy and trustworthiness. By addressing key data bottlenecks and seamlessly fusing synthetic and sparse real-world information, these methods pave the way for large-scale semantic 3D mapping, rapid disaster assessment, and next-generation digital twin technologies worldwide—all at a fraction of traditional costs.