Reproduce “SC-OmniGS: Self-Calibrating Omnidirectional Gaussian Splatting” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Huajian Huang, Yingshu Chen, Longwei Li, Hui Cheng, Tristan Braud, Yajie Zhao, Sai-Kit Yeung

SC-OMNIGS presents a self-calibrating framework for omnidirectional radiance field reconstruction from 360-degree images, leveraging 3D Gaussian Splatting. It jointly optimizes 3D Gaussians, camera poses, and a novel differentiable omnidirectional camera model to achieve high-fidelity scene representations robustly, even with noisy camera inputs or without pose priors.

Key insights

  • Directly operating on the spherical projection of 360-degree images, rather than converting to cube maps, maintains data integrity and efficiency for omnidirectional radiance field reconstruction.
  • Learning a generic, differentiable omnidirectional camera model significantly improves reconstruction quality by adapting to complex, real-world distortion patterns specific to 360-degree camera systems.
  • Applying a spherical weighting function in the photometric loss effectively addresses the unbalanced spatial resolution of equirectangular projections, ensuring spatially equivalent optimization across the entire sphere.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.