[1Yin-A] Poster 1

[1Yin-A-07] 4D Urban Dynamic Scene Reconstruction from Multiple Panoramic Videos

〇Hina Kogure¹, Kei Katsumata¹, Taiki Miyanishi^2,1, Komei Sugiura¹ (1. Keio University, 2. Graduate School of Engineering, The University of Tokyo)

Keywords:

4D Reconstruction,Gaussian Splatting,Urban Dynamics

We focus on reconstructing the dynamics of urban scenes as a unified 4D representation, using panoramic videos captured from multiple locations. In urban environments, physical and social constraints limit camera placement, resulting in sparse viewpoints and limited view overlap across locations. Under such conditions, existing methods designed for a single camera location often result in unstable optimization in 3D structure estimation. Our approach uses camera location information to align multi-view panoramic observations in a common world coordinate system while preserving geometric consistency. This enables dynamic structures observed from different viewpoints to be integrated into a unified coordinate frame. To evaluate our approach, we construct a novel 4D reconstruction benchmark based on panoramic videos captured from multiple locations in urban environments. Experimental results show that our method outperformed existing 4D reconstruction approaches designed for single-location observations on standard evaluation metrics.