ECCV 2026

UniTriSplat

A Unified 3D Gaussian Splatting Framework with Uniform Spherical Rasterization for Universal Cameras

Yipeng Zhu · Huajian Huang† · Tristan Braud · Sai-Kit Yeung

The Hong Kong University of Science and Technology  ·  Beijing Institute of Technology

† Corresponding author

Paper Supplementary arXiv Code HSSIM · Coming soon

Overview

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, including perspective, fisheye, and omnidirectional cameras. UniTriSplat is 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, it constructs a spherical sampling grid aligned with the angular resolution of input images and derives forward rendering and gradient propagation directly in the spherical radian domain. This enables uniform optimization behavior from narrow-FoV images to full 360-degree panoramas. A HEALPix-aware SSIM loss further respects spherical neighborhood structure. Extensive experiments across diverse camera models demonstrate improved cross-camera generalization while preserving geometric fidelity and rendering quality.

Demo video. The same spherical rasterizer supports perspective, fisheye, and omnidirectional rendering.

Method

Pipeline. Heterogeneous inputs are mapped to HEALPix, rasterized in spherical radian space, and supervised consistently.
More Method Details
Rasterization. Equal-area HEALPix pixels stabilize spherical sampling, depth compositing, and tile queries.
FoV support. Only the image-to-sphere map and visible FoV region change across camera models.
H-SSIM. Zone-aware kernels and boundary handling keep structural supervision consistent on HEALPix.

Results

Multi-FoV Evaluation

This evaluation compares UniTriSplat with camera-specific baselines across omnidirectional, fisheye, and perspective scenes with diverse FoVs.

Omnidirectional

Omnidirectional Scene 1

360Roam · FoV icon pending

Cross-camera Validation

Omnidirectional models are re-rendered through fisheye and perspective rasterizers to compare cross-projection consistency across camera models.

Ricoh360

Pillar

Reconstruction quality of each method on the omnidirectional scene

Additional Results
Ablation Study
Ablation study. RING scan, H-SSIM, and radian-space density control trade off quality and efficiency.
Cross-Resolution Evaluation
Cross-resolution evaluation. Additional 360Roam comparisons at multiple input resolutions.
FIORD Fisheye Inputs
FIORD fisheye inputs. Comparisons against OP43DGS, Fisheye-GS, and 3DGUT.
ScanNet++ Fisheye Inputs
ScanNet++ fisheye inputs. Additional indoor fisheye reconstruction comparisons.
Mip-NeRF 360 Perspective Inputs
Mip-NeRF 360 perspective inputs. Additional perspective reconstruction comparisons.
Vanilla 3DGS Scene: Render to Fisheye
Vanilla 3DGS Scene: Render to fisheye. Fisheye rendering of perspective-trained vanilla 3DGS scenes.
Vanilla 3DGS Scene: Render to Omnidirectional
Vanilla 3DGS Scene: Render to omnidirectional. Omnidirectional rendering of perspective-trained vanilla 3DGS scenes.

Citation

The official proceedings citation and BibTeX entry will be added when available.

Expanded research figure