Zixun Huang · Cho-Ying Wu · Yuliang Guo · Xinyu Huang · Liu Ren
Bosch Center for AI, Bosch Research North America
Check Project Page for More Visuals
3D Gaussian Splatting (3DGS) has rapidly become one of the most influential paradigms in neural rendering. It delivers impressive real-time performance while maintaining high visual fidelity, making it a strong alternative to NeRF-style volumetric methods. But there is a fundamental problem hiding beneath its success:
Splatting doesn't obey exactness in projective geometry.
The splatting approximation is usually harmless for narrow field-of-view (FoV) pinhole cameras. However, once we move to fisheye, omnidirectional, or generic camera models — especially those common in robotics and autonomous driving — the approximation error becomes significant.
This repository contains the official authors implementation associated with the ICLR 2026 paper "3DGEER: 3D Gaussian Rendering Made Exact and Efficient for Generic Cameras". The gsplat-geer OSS extension can be found here.
- Projective exactness + Real-time efficiency
- Compatibility with generic camera models (pinhole / fisheye) + Strong generalization to extreme FoV
- Adaptation to widely-used GS frameworks including
diff-gaussian-rasterization,gsplat,drivestudio
@misc{huang20263dgeer3dgaussianrendering,
title={3DGEER: 3D Gaussian Rendering Made Exact and Efficient for Generic Cameras},
author={Zixun Huang and Cho-Ying Wu and Yuliang Guo and Xinyu Huang and Liu Ren},
year={2026},
eprint={2505.24053},
archivePrefix={arXiv},
primaryClass={cs.GR},
url={https://arxiv.org/abs/2505.24053},
}- TBD:
drivestudio-geerandstormGaussian-geerwill be released here as well! - 2026-03-09:
gsplat-geerreleased here! - 2026-03-09: Code released! Can Gaussian rendering be both exact and fast without relying on lossy splatting? Check out 3DGEER!
- 2026-03-09: Code release approved. License updated. Requested admin to push code to BoschResearch.
- 2026-01-25: 3DGEER accepted to ICLR 2026, with an initial review of average 7 (top 1% score).
- 2025-05-29: Preprint released on Arxiv.
The full CUDA implementation can be found here: ./submodules/geer-rasterizer/.
-
Ray–Gaussian Integral (Forward & Backward): Analytical forward rendering and numerical stable backward gradient computation. (See paper Appendix C for the math.)
-
Particle Bounding Frustum: Efficient AABB for ray–particle association. (See paper Appendix D for the math.)
Following the 3dgs dependencies https://github.com/graphdeco-inria/gaussian-splatting to install the 3dgs environment, and then run the following command to replace the diff-gaussian-rasterization for using a geer-version CUDA rasterizer:
pip install ./submodules/geer-rasterizerSet you data path and 3dgeer codebase path in ./docker/init_my_docker.sh.
# Build up 3dgs environments for 3DGEER. Example:
bash ./docker/build.sh 4090
# Reset Docker on Terminal 1
bash ./docker/init_my_docker.sh
# If you modify algorithm upon our geer-rasterizer, inside docker container, recompile:
pip install --no-build-isolation ./submodules/geer-rasterizer# Enter Workspace for SIBR Viewer on Terminal 2
bash ./docker/run_my_docker.sh
# Inside docker container, run:
$sibr_rgOur framework follows the standard COLMAP data structure. For generic cameras (e.g., Fisheye), ensure your cameras.txt includes the specific intrinsic parameters. Link to detailed data format documentation.
Expected Directory Structure:
|_./data/scnt
|_datasets # e.g., download data into this folder
|_1d003b07bd
| |_colmap
| | |_images.txt
| | |_points3D.txt
| | |_cameras.txt
| | |_...
| |_nerfstudio
| | |_transforms.json
| |_resized_images
| |_000000.jpg
| |_000001.jpg
| |_...
|_e3ecd49e2b
|_...
To train 3DGEER on scannet++ data:
bash ./scripts/train_scnt.shfull training codes and scripts will be released soon.
To render high-quality images and compute PSNR/SSIM/LPIPS:
bash scripts/render_scnt.sh <SCENE_ID> <DATA_ROOT> <CKPT_DIR> <MODE>
bash scripts/eval_scnt.sh <SCENE_ID> <DATA_ROOT> <CKPT_DIR> <MODE>Arguments:
SCENE_ID : scene name (e.g. steakhouse, 1d003b07bd/dslr)
DATA_ROOT : root directory of the formatted dataset
CKPT_DIR : directory containing the trained model checkpoint
MODE : rendering backend, (BEAP, KB or PH)
Set
DIST_SCALINGas 0 in the shell to render EQ under KB mode; Enlarge the value ofFOCAL_SCALINGto test extreme large FoV; For fair comparison, we recommend evaluating withBEAPmode, which ensures consistent metric computation across different rendering backends.
Example: See examples in detailed train and eval documentation.
Please ensure that the corresponding ground truth is used. For example, evaluating extreme KB images using the original KB images as ground truth is invalid due to mismatched distortion parameters.
You can download the pre-trained checkpoints for the scenes shown on our project webpage:
- ScanNet++: Kitchen, Lab, Officeroom, Bedroom
- ZipNeRF: Alameda, Berlin, London, NYC
- Aria: Livingroom, Steakhouse, Garden
- Tank and Temples: Train, Truck
- Customized Parking: Bosch Center
Download from HuggingFace: https://huggingface.co/datasets/ZixunH/3DGEER_ckpt
3DGEER supports the opensource community with gsplat integration.
Check out our gsplat-geer branch for details.
Feel free to drop a pull request whenever!
👀Visuals (More)
Side-by-Side Comparison on Close-Up Parking Data: 3DGEER's PBF association (Right Col.) has less popping issues (First Row) and no grid-line artifacts (Second Row) compared w/ UT (Left Col.).
3DGEER is released under the AGPL-3.0 License. See the LICENSE file for details. This project is built upon 3D Gaussian Splatting by Inria. We thank the authors for their excellent open-source work. The original license and copyright notice are included in this repository, see the file 3dgs-license.txt.