Reproduce “3D Gaussian Splatting with Fisheye Images: Field of View Analysis and Depth-Based Initialization” 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

Ulas Gunes, Matias Turkulainen, Mikhail Silaev, Juho Kannala, Esa Rahtu

This work provides the first systematic evaluation of 3D Gaussian Splatting methods adapted for fisheye cameras on real-world images with fields of view exceeding 180°, analyzing performance across varying FoVs and with monocular depth-based initialization. It identifies a 160° FoV as an optimal balance for reconstruction quality and demonstrates that UniK3D-based initialization reduces preprocessing time for Fisheye-GS from hours to seconds while maintaining comparable quality.

Key insights

  • The 160° FoV offers a balanced trade-off between scene coverage and mitigation of adverse effects from peripheral distortion for fisheye 3DGS.
  • 3DGUT is generally more effective in compact, distortion-heavy indoor scenes, while Fisheye-GS demonstrates greater stability in larger, multi-view environments.
  • Monocular depth estimation via UniK3D provides a practical and significantly faster alternative to SfM for initializing Fisheye-GS, even without explicit training on real extreme FoV data.

Summary by alphaXiv.

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