Transitioning from Calibrated to Uncalibrated Sparse-View 3D Reconstruction via High-Redundancy DUSt3R Pipelines

Recent advancements in 3D reconstruction, specifi- cally 2D Gaussian Splatting (2DGS), have enabled high-fidelity surface modeling with rapid optimization. However, state-of-the- art sparse-view frameworks like Sparse2DGS remain tethered to laboratory conditions due to their strict requirement for ground- truth camera parameters. This dependency limits their utility in practical, uncalibrated ”in-the-wild” scenarios where camera metrics are unknown. In this paper, we investigate the feasibility of uncalibrated sparse-view reconstruction by evaluating two distinct methodologies. First, we test a Hybrid SfM-DUSt3R pipeline that utilizes deep-learning-based pose estimates to drive traditional COLMAP Multi-View Stereo (MVS). We find this approach to be fundamentally fragile, often resulting in catas- trophic sparsity (fewer than 15 points) due to the intolerance of MVS algorithms to minor pose inaccuracies. To overcome this, we propose an Autonomous High-Redundancy Pipeline that bypasses traditional Structure from Motion (SfM) entirely. By increasing view redundancy (10+ images) and implementing a post-inference refinement suite—comprising Projective Masking and Statistical Outlier Removal—we successfully initialize 2D Gaussian primitives with over 2.3 million dense points. Our results demonstrate that this redundancy-driven approach pro- vides a robust, watertight reconstruction that is significantly more reliable for real-world applications than hybrid-calibration methods.