Search papers, labs, and topics across Lattice.
This paper introduces SalientGS, a novel pipeline that integrates Structure-from-Motion (SfM) with 3D Gaussian Splatting (3DGS) through an importance-guided MCMC Gaussian allocation method. By leveraging multi-view residuals to inform a smooth sampling distribution, the approach efficiently reallocates computational resources to underfit regions, enhancing reconstruction quality without compromising the stochastic gradient Langevin dynamics. SalientGS achieves rapid end-to-end 3D scene reconstruction in just 15 minutes, setting a new benchmark for perceptual quality in this domain.
SalientGS reallocates computational resources to underfit areas, achieving high-quality 3D reconstructions in just 15 minutes.
Reconstructing 3D scenes from unordered images remains bottlenecked by expensive Structure-from-Motion (SfM) preprocessing and frozen pose interfaces. We present SalientGS, a unified SfM-to-3D Gaussian Splatting (3DGS) pipeline. Its central contribution is importance-guided Markov Chain Monte Carlo (MCMC) Gaussian allocation, which aggregates multi-view residuals into per-Gaussian underfit and redundancy signals. These signals define a smooth importance-weighted sampling distribution that biases both birth and relocation toward underfit regions. This reallocates capacity from well-fit areas without altering the underlying stochastic gradient Langevin dynamics (SGLD). SalientGS achieves end-to-end reconstruction in 15 minutes with state-of-the-art perceptual quality. The supplementary material provides dedicated sections for Per-Scene Qualitative Comparisons and Per-Image Learned Perceptual Image Patch Similarity (LPIPS) Analysis, including failure cases. Code and evaluation scripts are available at https://github.com/Six-Bit-TX/SalientGS.
