--- license: cc-by-nc-4.0 pretty_name: GOFish scenes size_categories: - n<1K task_categories: - image-to-3d tags: - gaussian-splatting - 3d-reconstruction - mesh - fisheye --- # GOFish: Full Field-of-View Gaussian Splatting and Mesh Extraction [Allen Tu](https://tuallen.github.io/)1,2, [Meher Gitika Karumuri](https://scholar.google.com/citations?user=R14csHAAAAAJ)1, [Shifan Zhu](https://shifanzhu.github.io/)1,3, [Hailing Wang](https://scholar.google.com/citations?user=oKoYjqUAAAAJ)1,4, [Yen-Liang Lin](https://sites.google.com/view/yenlianglin)1, [David Held](https://davheld.github.io/)1,5, [Matthias Zwicker](https://www.cs.umd.edu/~zwicker/)2, [Tom Goldstein](https://www.cs.umd.edu/~tomg/)2 1Amazon, 2University of Maryland, College Park, 3University of Massachusetts Amherst, 4Northeastern University, 5Carnegie Mellon University [Project page](https://3dgofish.github.io/) | [Paper](https://arxiv.org/abs/2610.37818) | [Code](https://github.com/tuallen/GOFish) ![GOFish reconstructs FIORD Hall_In as a Gaussian Splatting scene and aligned mesh over its full 200° field of view; 3DGUT + PPISP stops at 160°.](assets/teaser.jpg) ## Scenes: FIORD and FullCircle Each folder `_/` holds one scene reconstructed over the full fisheye field of view (FIORD 200 degrees, FullCircle about 205 degrees per lens), as evaluated in the GOFish paper. | Folder | Contents | |---|---| | `gaussians/` | GOFish (Ours): the Gaussian Splatting scene trained without SOF's geometry losses. Used for all rendering results. | | `gaussians_geo/` | GOFish + Geometry Objective: the same pipeline with SOF's geometry losses. | | `mesh/` | The + Geometry Objective mesh used in the paper's mesh results, as binary PLY and Draco-compressed GLB. | | `cameras.json` | Calibration (OpenCV fisheye model) and every train/test camera pose. | | `lidar_alignment.json` | FIORD geometry scenes only: the similarity transform into the LiDAR ground-truth frame used for the paper's mesh scores. Corridor_Out has none because its SfM-LiDAR registration is invalid, so the paper evaluates it for rendering only. | ## Download Download one scene, or omit `--include` for all of them: ``` hf download AllenTu/GOFish-scenes --repo-type dataset --include "FIORD_Kitchen_In/*" --local-dir GOFish-scenes ``` ## Formats - `gaussians*/point_cloud.ply`: standard 3D Gaussian Splatting (INRIA) layout: `x y z nx ny nz f_dc_0..2 f_rest_0..44 opacity scale_0..2 rot_0..3`. Opacity is a logit (apply sigmoid), scales are log (apply exp), rotation is a quaternion w,x,y,z, colors are degree-3 spherical harmonics with channel-major `f_rest`. The GOFish 3-D Mip filter is baked in: each stored scale is `sqrt(scale^2 + filter_3D^2)`, so the file is exactly the stored scene GOFish renders and meshes. Opacity is not rescaled, matching GOFish. - `gaussians*/appearance.json`: the PPISP sensor model and GOFish's wide-angle attenuation W. This covers per-image exposure and color, per-camera vignetting, the camera response function, the radial profile and flat field, and the novel-view controller weights. Tensors are flattened row-major with their shapes. Generic viewers ignore this file, so they show the scene's radiance without the camera's exposure and lens falloff. - `mesh/mesh.ply`: the mesh exactly as evaluated, byte-identical to the paper's file. `mesh/mesh.glb` is the same mesh with Draco compression (KHR_draco_mesh_compression, 16-bit positions) for web viewers. - `cameras.json`: COLMAP convention, `x_cam = R(q) x_world + t`, with `camera_to_world` given as its inverse. Intrinsics are at the stored image resolution. GOFish trained on images downsampled by 4, so divide fx, fy, cx and cy by 4 for the training resolution. Gaussians, mesh and cameras share one world frame, the training COLMAP registration, with no normalization. ## Viewing - Gaussians: SuperSplat (https://superspl.at/editor), gsplat, Nerfstudio or any 3DGS PLY viewer. These render with a pinhole projection. GOFish's full-FoV fisheye rendering needs the GOFish/3DGUT renderer with `cameras.json` and `appearance.json`. - Mesh: MeshLab, Blender or Open3D for `mesh.ply`; three.js, Babylon.js or a glTF viewer with Draco support for `mesh.glb`. `MANIFEST.json` in each folder lists every file with its size and sha256, and each scene's `README.md` gives its counts and provenance. ## Citation If you use these scenes, please cite GOFish and the source dataset. ```bibtex @misc{Tu2026GOFish, author = {Tu, Allen and Karumuri, Meher Gitika and Zhu, Shifan and Wang, Hailing and Lin, Yen-Liang and Held, David and Zwicker, Matthias and Goldstein, Tom}, title = {{GOFish}: Full Field-of-View Gaussian Splatting and Mesh Extraction}, year = {2026}, eprint = {2610.37818}, archivePrefix = {arXiv}, primaryClass = {cs.CV}, url = {https://3dgofish.github.io/} } @inproceedings{gunes2025fiord, author = {Gunes, Ulas and Turkulainen, Matias and Ren, Xuqian and Solin, Arno and Kannala, Juho and Rahtu, Esa}, title = {{FIORD}: A Fisheye Indoor-Outdoor Dataset with {LIDAR} Ground Truth for {3D} Scene Reconstruction and Benchmarking}, booktitle = {Proceedings of the Scandinavian Conference on Image Analysis (SCIA)}, pages = {3--17}, year = {2025} } @article{foroutan2026fullcircle, author = {Foroutan, Yalda and Oztas, Ipek and Rebain, Daniel and Dundar, Aysegul and Yi, Kwang Moo and Goli, Lily and Tagliasacchi, Andrea}, title = {{FullCircle}: Effortless {3D} Reconstruction from Casual 360$^\circ$ Captures}, journal = {arXiv preprint arXiv:2603.22572}, year = {2026} } ``` ## License These reconstructions are released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/), matching the GOFish code release. They are derived from the following datasets, whose licenses and attribution terms also apply: - **FIORD** (Gunes et al., SCIA 2025), [Zenodo](https://zenodo.org/communities/fiord), CC BY 4.0. - **FullCircle** (Foroutan et al., 2026), [Hugging Face](https://huggingface.co/datasets/youlenda/FullCircle), Apache 2.0. ## Acknowledgements This work was completed during an internship at [Amazon](https://www.amazon.science/). We thank [Ziyang Long](https://ziyanglong.cn/), [Mellon M. Zhang](https://meilongzhang.github.io/), [Chenqi Li](https://scholar.google.com/citations?user=u_lYTIwAAAAJ), [Sarosij Bose](https://sarosijbose.github.io/), [Srikumar Sastry](https://vishu26.github.io/), [Amin Kheradmand](https://scholar.google.com/citations?user=5DJL154AAAAJ), [David Stafford](https://www.linkedin.com/in/davidwstafford), [Yongqing Ye](https://www.linkedin.com/in/avatarye), [Huixian Gong](https://www.linkedin.com/in/huixiangong), [Yun He](https://yunhe20.github.io/), [Zeping Liu](https://zpl99.github.io/), [Haiyang Ying](https://oceanying.github.io), [Alex Hanson](https://www.cs.umd.edu/~hanson/), [Chengyue Huang](https://chengyuehuang511.github.io/), and Chen Xu for helpful discussions and feedback on this work.