CoverPruneGS: Coverage-Preserving Structured Pruning for Hierarchical 3D Gaussian Splatting from Sparse-View Monocular Videos
Abstract
Reconstructing a complete yet compact 3DGS from sparse-view monocular long videos is challenging: hierarchical training with VFI can improve coverage, yet correlated pseudo views and repeated merging tend to accumulate near-duplicate Gaussians and exacerbate co-adaptation. To address this, we propose CoverPruneGS, a coverage-preserving structured pruning framework tailored for hierarchical merging with VFI-augmented supervision, which performs coarse-to-fine pruning via voxel-based local diversity selection and GT-guided lazy refinement with randomized dropout rendering. To make refinement reliable, we introduce a footprint-aware CUDA attribution that aggregates GT-aligned error degradation over Gaussian-influenced pixels in a manner consistent with alpha compositing, yielding faithful per-Gaussian scores for quantile-based rescue. Experiments on multiple datasets demonstrate that CoverPruneGS substantially reduces the Gaussian count by 56.8\% and accelerates inference while maintaining or improving novel view synthesis quality.