GOFish: A Method to Obtain Both a Beyond-180° 3DGS and a Collision Mesh from Fisheye Images

What happened

GOFish (Full Field-of-View Gaussian Splatting and Mesh Extraction) is a research project that builds 3D Gaussian Splatting (3DGS) and a mesh from fisheye images. The authors are affiliated with Amazon, the University of Maryland, Carnegie Mellon University, and others. The project page has links to the paper, arXiv, code, models, and a thread.

How it works

As an existing fisheye-capable method, the project page cites 3DGUT. It describes two problems with 3DGUT. First, it truncates the field of view to 160° and sorts by depth along the optical axis, so the order is reversed beyond 90°. Second, it approximates tile coverage, which causes ring-shaped gaps and tile-like artifacts. GOFish determines the valid image region from calibration information. It renders every Gaussian reached by a valid ray, and at each pixel composites from front to back using ray distance. For densification (how Gaussians are added), it measures gradients in image space, and says a single threshold is enough to recover detail across the full field of view. Calibrated Mip Filtering uses the camera Jacobian to separate a view-independent scene-side filter from each view's pixel footprint. The pixel footprint is applied only at render time. According to the authors, this keeps per-view blur out of the scene and the mesh, and also keeps the number of Gaussians down. Wide-Angle Appearance Modeling extends the existing PPISP. It combines a monotonic radial profile with a low-resolution flat field, and treats the steep falloff at the periphery as a camera characteristic. The authors say this prevents peripheral darkness from being baked into the Gaussians' colors and carried over into the mesh.

By the numbers

Why it matters

Robots, autonomous vehicles, and 360° cameras often use fisheye lenses. For workflows that build reconstructed real-world environments from captured footage, being able to use the full field of view out to the edges could be an advantage. However, whether it works in real-world settings depends on what code and models are released and on testing with your own data.

How to try it

1. Check the Code and Models links on the project page and look into their release status and license.
2. Because the method requires fisheye lens calibration values, check whether you can obtain the intrinsic parameters of your own camera.

Caveats and open questions

• All figures are values announced by the authors on the project page; third-party verification is not confirmed in the sources.
• The positioning as a "first" method is also the authors' claim.
• Processing time, GPU requirements, real-time rendering frame rate, and training time have no specific values in the sources and are unconfirmed.
• Links to code and models exist, but the license and release status are not confirmed in the sources.
• Quantitative values for ScanNet++ and Tanks & Temples, and detailed table values, cannot be read from the body text of the sources.
• The sources give the arXiv number as "2610.37818", but the paper itself has not been checked.

Glossary

3D Gaussian Splatting (3DGS)

A method that represents a scene with many 3D Gaussians and renders it quickly.

Fisheye lens

A lens that can capture a wide field of view of around 180° or more.

PSNR

A metric of image reconstruction accuracy; higher values mean the result is closer to the original image.

F-score

A metric that evaluates the accuracy of shapes such as meshes using precision and recall.

Digital twin

A digital reproduction of a real place or object in a form usable for simulation and similar purposes.

Collision mesh

A surface geometry used by robots and similar systems to detect contact with objects.

Sources

Primary sources (publisher, paper, repository)

• GOFish: Full Field-of-View Gaussian Splatting and Mesh ExtractionGOFish プロジェクトページ（3dgofish.github.io）・en

The X post that surfaced this story

• View @tuallen_'s post on Xx.com

This article was generated in Japanese by AI from the sources listed (primarily primary sources) and published automatically, without prior human review, after automated checks such as source matching (how we use AI). Figures and specifications are as stated by the publishers and have not been independently verified by us. Please consult the original sources. To report an error, use our contact form; we will review it promptly and correct it. This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails. Japanese original