CoRe-GS: Coarse-to-Refined Gaussian Splatting with Semantic Object Focus

Abstract

Fast and efficient photorealistic 3D reconstruction with (semantic) Gaussian Splatting (GS) is crucial for time-critical robotic perception and navigation, where robots may need to rapidly reconstruct and inspect specific points of interest (POIs), such as in search-and-rescue scenarios. Existing semantic GS methods prioritize reconstruction and segmentation quality while uniformly optimizing entire scenes, often overlooking runtime constraints crucial for real-world deployment. Moreover, object extraction often suffers from floaters and incomplete reconstructions, while post-optimization approaches often produce incomplete POIs. We propose CoRe-GS, a coarse-to-refined, task-driven approach that first creates a segmentation-ready GS representation through lightweight late-stage semantic refinement, then selectively refines only the views and Gaussians associated with the POI. A color-based filtering strategy further suppresses segmentation-induced floaters. On LERF-Mask, our representation achieves competitive mIoU with 6-8x fewer optimization steps. On NeRDS 360 and SCRREAM, CoRe-GS reduces time-to-POI by 91-95% over full-scene semantic GS while improving reconstruction quality and reducing floaters. Across all benchmarks, it is 44-76% faster than post-optimization object extraction, demonstrating that task-aware refinement can deliver efficient, high-quality POI reconstruction for time-critical robotic perception and operation.