Probabilistic 4D Scene Reconstruction from Monocular Videos

Speaker: Professor Bohyung Han
Department of Electrical and Computer Engineering
Seoul National University (SNU)
Korea

Title: Probabilistic 4D Scene Reconstruction from Monocular Videos

Date: Monday, 9 February 2026

Time: 4:00pm - 5:00pm

Venue: Lecture Theater F
(Leung Yat Sing Lecture Theater), near lift 25/26, HKUST

Abstract:

Reconstructing dynamic 4D scenes from casually captured monocular videos is a challenging problem due to sparse observations, fast motion, occlusions, and the lack of reliable priors in real-world settings. In this talk, I present our recent efforts to advance 4D Gaussian Splatting (4DGS) by explicitly modeling uncertainty and dynamics in both geometric and probabilistic frameworks. I first introduce an uncertainty-aware regularization strategy that identifies poorly observed regions and selectively imposes additional priors, together with a depth- and scene-flow-based densification scheme that robustly initializes Gaussian primitives in fast-moving regions where structure-from-motion fails. Building on these insights, I then present GP-4DGS, a probabilistic formulation that models the motion of 4D Gaussian primitives using variational Gaussian Processes, enabling flexible, data-adaptive motion modeling, principled uncertainty estimation, and temporal extrapolation beyond observed frames. Finally, I briefly discuss our ongoing effort toward physics-aware evaluation of dynamic novel view synthesis, including the construction of a dataset designed to support future research on physically grounded 4D scene modeling.

Biography:

Bohyung Han is a Professor in the Department of Electrical and Computer Engineering at Seoul National University (SNU), Korea. Prior to joining SNU in 2018, he was an Associate Professor in the Department of Computer Science and Engineering at Pohang University of Science and Technology (POSTECH). He has also gained extensive industrial research experience as a Visiting Research Scientist at Google DeepMind, Google Research, and Snap Research in the United States. He received his Ph.D. in Computer Science from the University of Maryland, College Park, in 2005. His research interests span computer vision and machine learning. Prof. Han has made significant contributions to the research community through numerous leadership roles, including Program Chair for ICCV 2025 and BMVC 2026, Technical Program Committee Vice-Chair for ICASSP 2024, and General Chair for ACCV 2022. He has also served as a Senior Area Chair for CVPR, NeurIPS, ICLR, and ICML, and is currently an Associate Editor of the IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI).