GCS-Splat: Geometric Consistency Score-Driven Active Reconstruction with 3D Gaussian Splatting
Abstract
Existing active reconstruction methods often suffer from structural holes in 3D Gaussian Splatting (3DGS) due to the over-expansion and geometric distortion of Gaussians. To address this issue, we propose GCS-Splat, an active reconstruction framework focused on the intrinsic geometric quality of Gaussians. Our approach tightly couples intrinsic Gaussians geometric repair with active motion viewpoint planning. At the core of this framework is the Geometric Consistency Score (GCS), which quantifies Gaussian distortion to evaluate local reconstruction uncertainty. Guided by the GCS, we perform probabilistic resampling of historical frames based on geometric correlations, and direct the pruning, splitting, and cloning of Gaussians. For motion viewpoint planning, we design a two-stage coverage-repair planner: it first rapidly achieves spatial coverage through autonomous exploration of global frontiers, and subsequently conducts multi-view repair and path planning targeted at geometrically defective regions under GCS guidance. Furthermore, all generated trajectories are smoothed using B-splines to ensure superior kinematic performance. Extensive experiments on the Replica and MP3D datasets demonstrate that our method achieves favorable performance in both geometric accuracy and reconstruction quality, unifying efficient global exploration with high-fidelity reconstruction.