Home | Publications | XJC+25

CSG-Fusion: Consistent Sparse-View Gaussian Splatting via Matching-Based Fusion

MCML Authors

Weirong Chen

→ Group Daniel Cremers
Computer Vision & Artificial Intelligence

Daniel Cremers

Prof. Dr.

Director

Computer Vision & Artificial Intelligence

Abstract

Recent developments in Gaussian splatting have enabled high-fidelity 3D reconstruction from multi-view images, but pixel-aligned methods such as MASt3R often produce redundant primitives and inconsistent geometry under few-view settings. We propose CSG-Fusion, a feed-forward framework that mindfully integrates pixel-aligned pointmap to reduce redundant primitives and produce compact and consistent 3D structures. Our approach leverages a matching prior with spatial thresholds to prune overlapping Gaussians, forming a coherent base 3D model, and then applies a mask-based feature aggregation module to merge local features and improve photometric consistency with fewer primitives. To enforce cross-view agreement after fusion, we further incorporate context-view supervision to align appearance and geometry across perspectives. Experiments on the large-scale ScanNet++ and object-level DTU benchmarks demonstrate both the efficiency and generalization of our method. Compared to the leading pose-known and pose-free approaches, our method achieves higher rendering quality with substantially fewer Gaussians.

inproceedings

E2E3D @ICCV 2025

Workshop on End-to-End 3D Learning at the IEEE/CVF International Conference on Computer Vision. Honolulu, Hawai'i, Oct 19-23, 2025. Best Paper Award. To be published. Preprint available.