Reconstructing Reality over Time: From Drone Capture to Timelapse Gaussian Splatting

초록

We present a system for reconstructing and visualizing time-varying 3D scenes from multi-day drone imagery using Gaussian Splatting (GS). While existing Gaussian Splatting methods excel at static or short-term dynamic scenes, long-term outdoor capture presents new challenges due to temporal inconsistencies and computational overhead when processing thousands of images across days. We introduce an automated alignment pipeline combining GPS-guided spatial matching, geo-registration, and multi-stage ICP refinement to enable temporally consistent reconstruction without manual intervention or physical markers. Applied to our 30-day construction site dataset, our method demonstrates photorealistic timelapse reconstruction and interactive novel view rendering through a web-based client. This case study shows how 3DGS can be extended to real-world long-term scene monitoring applications. © 2025 IEEE.

키워드