3dgeer

Zixun Huang · Cho-Ying Wu · Yuliang Guo · Xinyu Huang · Liu Ren

Bosch Center for AI, Bosch Research North America

🧐Overview

😺Key Features

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.

🎉News

• TBD: stormGaussian-geer will be released here as well.
• 2026-06-25: drivestudio-geer is released under our gsplat-geer branch, including a customized dynamic, temporal viewer!
• 2026-03-19: 3DGEER now supports dynamic outdoor scene rendering under wide-FoV fisheye cameras with the integration into DriveStudio.
• 2026-03-17: SIBR_remoteGaussian_app is adapted to our work as an interactive viewer for training and trained checkpoints. Try the BEAP mode in the viewer; Pinhole and Fisheye modes are supported as well.
• 2026-03-09: gsplat-geer released 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.

📷3DGEER-CUDA-Rasterizer

The full CUDA implementation can be found here: ./submodules/geer-rasterizer/.

Key Insight 1: Fixing the Math Behind Gaussian Projection

• Ray–Gaussian Integral (Forward & Backward): Analytical forward rendering and numerical stable backward gradient computation. (See paper Appendix C for the math.)

Key Insight 2: Fixing the Math Behind Gaussian Association

• Particle Bounding Frustum: Exact and minimal boundary geometry for ray–particle association. (See paper Appendix D for the math.)

Key Insight 3: Optimizing the Ray Distribution Behind Pixelwise Color Supervision

• Bipolar Equiangular Projection: Maintains uniform ray sampling across arbitrary fields of view, thereby providing stable, FoV-invariant supervision for radiance field training.

🔧Dependency

Conda Based Installation

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-rasterizer

Docker Configuration (Recommend)

Set 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

🔧Interactive Viewer Setup

SIBR Viewer Configuration with Docker

⚠️ Important Notice on Visualization: SIBR_gaussianViewer_app is currently not supported for Gaussian Exact and Efficient Rendering (GEER). Please do not use:

  $sibr_gv -m "./output/scnt/<SCENE_ID>/dslr"

as it invokes the vanilla 3D Gaussian Splatting rasterizer for offline rendering from checkpoints. This leads to invalid results, since GEER-trained (ray-based) scenes are incompatible with splatting-based rendering.

✅ Recommended Alternatives:

• During training, use SIBR_remoteGaussian_app, which connects via port and calls our modified GEER rasterizer.

# Enter Workspace for SIBR Viewer on Terminal 2
bash ./docker/run_my_docker.sh
# Inside docker container, run:
$sibr_rg

• To use SIBR_remoteGaussian_app from checkpoints, first run the following on Terminal 1.

bash scripts/visualize.sh <SCENE_ID> <DATA_ROOT> <CKPT_DIR> <MODE: BEAP, PH or KB> # Example: 1d003b07bd/dslr data/scnt/datasets ckpt/scnt KB
# Then (on Terminal 2) launch:
$sibr_rg

• For offline visualization, We recommend using our gsplat-geer implementation, built on top of: https://github.com/nerfstudio-project/gsplat/blob/main/docs/3dgut.md

Note: the mismatched culling issue in UT is resolved using our PBF-based fix.

🏃Quick Start

1. Data Preparation

Our 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
        |_...

2. Training 3DGEER

To train 3DGEER on ScanNet++ data in BEAP mode (default, recommended):

bash ./scripts/train.sh

Key training arguments:

Argument	Description	Default
-s / --source_path	Path to the scene dataset directory	—
-m / --model_path	Output directory for checkpoints and logs	./output/
--render_model	Training projection mode: BEAP, KB, EQ, or PH	BEAP
--sample_step	Ray sampling interval in radians (BEAP/KB)	—
--fov_mod	FoV scale factor applied during training (BEAP mode)	—
--mask_path	Path to validity mask PNG	—
--raymap_path	Path to per-pixel ray-direction map .npy (KB/EQ mode)	—
--focal_scaling	Scale factor for focal length (KB/PH modes)	1.0
--distortion_scaling	Scale factor for distortion coefficients (KB mode; 0 → EQ)	1.0
--mirror_shift	Mirror-model shift ξ for omnidirectional mapping (KB)	0.0
--iterations	Total training iterations	30000

See detailed training documentation and examples for BEAP, KB, and PH mode commands.

3. Rendering & Evaluation

To render high-quality images and compute PSNR/SSIM/LPIPS:

bash scripts/render.sh <SCENE_ID> <DATA_ROOT> <CKPT_DIR> <MODE>
bash scripts/eval.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_SCALING as 0 in the shell to render EQ under KB mode; Enlarge the value of FOCAL_SCALING to test extreme large FoV; For fair comparison, we recommend evaluating with BEAP mode, 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.

4. Available Checkpoints

You can download our trained checkpoints for the following open dataset:

• ScanNet++: Kitchen, Lab, Officeroom, Bedroom, Storage
• ZipNeRF: Alameda, Berlin, London, NYC
• Aria: Livingroom, Steakhouse, Garden
• Tank and Temples: Train, Truck

🙏Special Extension

⛽️Contributing

Feel free to drop a pull request whenever!

👀Visuals (More)

💡License

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.