UniTriSplat: A Unified 3D Gaussian Splatting Framework with Uniform Spherical Rasterization for Universal Cameras
2026-06-29 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors found that existing 3D Gaussian Splatting methods work differently depending on the camera type, causing inconsistent results. They created UniTriSplat, which works for all kinds of cameras by projecting the 3D data onto a sphere using a special grid called HEALPix. This approach makes the rendering and optimization smoother for all camera views, from normal lenses to full 360-degree ones. They also added a new way to compare images that respects the spherical shape, improving the final image quality. Tests showed their method works well across many camera types while keeping accurate shapes and good images.
3D Gaussian SplattingHEALPixunit spheresolid-angle samplingperspective camerafisheye cameraomnidirectional cameraspherical renderingSSIM losscross-camera generalization
Authors
Yipeng Zhu, Huajian Huang, Tristan Braud, Sai-Kit Yeung
Abstract
Existing 3D Gaussian Splatting (3DGS) frameworks rely on camera-specific rasterization, suffering from inconsistent solid-angle sampling and degraded performance across heterogeneous camera models (e.g., perspective, fisheye, omnidirectional). To address this limitation, we propose UniTriSplat, a unified 3DGS framework for universal cameras that reformulates Gaussian splatting on the unit sphere via HEALPix discretization. Leveraging the equal-area property of HEALPix, we construct a spherical sampling grid aligned with the angular resolution of input images. We derive the forward rendering and gradient propagation of Gaussians directly in the spherical radian domain, yielding uniform optimization behavior from narrow-FoV images to full 360-degree panoramas. To enhance perceptual reconstruction quality, we additionally introduce a HEALPix-aware SSIM loss that respects spherical neighborhood structure. Extensive experiments across diverse camera models demonstrate that UniTriSplat consistently improves cross-camera generalization while preserving geometric fidelity and rendering quality.