3DGS Viewer Logo
3DGS Viewer
Back to guide
English中文

Research Paper

CylinderSplat: 3D Gaussian Splatting with Cylindrical Triplanes for Panoramic Novel View Synthesis

A 2026 panoramic 3DGS paper that replaces Cartesian triplanes with cylindrical triplanes and combines pixel and volume branches for sparse 360-degree views.

March 2026Panoramic 3DGSarXiv:2603.05882
arXiv Paper->

Detailed Reading

CylinderSplat is important because panoramic capture is not just wide-angle capture with more pixels. A 360-degree image wraps around the viewer, has strong radial structure, and often follows indoor Manhattan-world layouts with vertical walls and horizontal floors. Cartesian triplanes can distort that geometry, especially when views are sparse and occlusions are large.

The central representation is the cylindrical triplane. Instead of decomposing a 3D volume along ordinary Cartesian planes, it uses coordinates better aligned with radius, angle, and height. This lets features follow the natural layout of panoramic scenes and reduces distortion compared with forcing a sphere or room into a Cartesian grid.

The architecture has two complementary branches. A pixel-based branch handles well-observed regions where the input panorama directly provides strong evidence. A volume-based branch uses the cylindrical triplane to infer occluded or sparsely observed regions, giving the model a structured space for completion.

The output is a feed-forward 3DGS representation, so the method does not require lengthy per-scene optimization in the same way as classic 3DGS. It can support variable numbers of input panoramas, from a single panorama to multiple views. That flexibility is crucial for real capture workflows where users may not follow an ideal scanning path.

Algorithmically, the paper shows that the intermediate feature geometry matters as much as the final Gaussian renderer. If the network reasons in a coordinate system aligned with the data, the predicted Gaussians become more geometrically accurate. This is a useful lesson for any generalizable 3DGS method that targets a special camera model.

The limitation is that panoramic priors are domain-specific. CylinderSplat is strongest when scenes resemble the assumptions behind cylindrical coordinates and indoor layout regularities. It is valuable for VR, real estate, and indoor capture, but less obviously general for arbitrary object scans or non-Manhattan outdoor scenes.

What The Paper Does

CylinderSplat targets panoramic novel view synthesis, where sparse 360-degree inputs make cost-volume refinement and Cartesian triplanes poorly matched to scene geometry.

It introduces cylindrical triplanes aligned with panoramic structure and a dual-branch architecture that combines visible-region reconstruction with occlusion completion.

Core Ideas

  • Introduces cylindrical triplanes for 360-degree scene geometry.
  • Uses a dual-branch design for observed pixels and occluded-volume completion.
  • Handles a variable number of panorama inputs.
  • Targets feed-forward panoramic novel view synthesis with Gaussian output.

Why It Matters

  • It adapts 3DGS to panoramic capture, an important format for VR and indoor spaces.
  • It shows how camera geometry should influence learned 3DGS representations.
  • It helps bridge single/multi-panorama capture and real-time splat rendering.

Read This If

  • You work with 360-degree cameras or panoramic scene capture.
  • You are designing feed-forward 3DGS models for special camera types.
  • You want an example of representation geometry improving Gaussian prediction.

Limitations And Caveats

  • The cylindrical prior may not suit arbitrary scenes or object-centric captures.
  • Occluded regions remain learned guesses when input views are sparse.
  • Feed-forward quality depends on training data and camera-domain match.