ConeGS: Error-Guided Densification Using Pixel Cones for Improved Reconstruction with Fewer Primitives

Bartłomiej Baranowski · Stefano Esposito · Patricia Gschoßmann · Anpei Chen · Andreas Geiger

🛠️ Installation

The code was tested on Python 3.11 with PyTorch 2.5.1 with CUDA Toolkit 12.1 and 11.8.

Installed CUDA Toolkit is required. To run the code install the following packages:

git clone https://github.com/baranowskibrt/conegs.git --recursive

conda create -n conegs python=3.11
conda install pytorch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 pytorch-cuda=12.1 -c pytorch -c nvidia
pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch
pip install git+https://github.com/nerfstudio-project/nerfacc.git
pip install submodules/simple-knn
pip install submodules/diff-gaussian-rasterization

pip install -r requirements.txt

📝 Usage

Single scene

For a particular scene run (Mip-NeRF360's bicycle scene as an example) with a specified budget run:

python train.py --config-name defaults.yaml gaussian_model.source_path=../scenes_mipnerf scene_name=bicycle run_name=benchmark optimization.max_points=100000 gaussian_model.images=images_4 

And without budget:

python train.py --config-name defaults.yaml gaussian_model.source_path=../scenes_mipnerf scene_name=bicycle run_name=benchmark optimization.max_points=100000 gaussian_model.images=images_4 

Benchmark

For all benchmarks on a specific budget adjust the max points (primitives), and run:

python full_eval.py --mipnerf360 ../scenes_mipnerf --tanksandtemples ../scenes_tt --deepblending ../scenes_db --ommo ../scenes --output_path benchmarks --common_args " optimization.max_points=100000"

Or for unbudget case:

python full_eval.py --mipnerf360 ../scenes/ --tanksandtemples ../scenes/ --deepblending ../scenes --ommo ../scenes --output_path benchmarks --common_args " optimization.max_points=0 optimization.gaussian_percentage_increase=0.02"

🔄 Pipeline

🛠️ Data

Datasets

We use the original release of every dataset:

• MipNeRF360 , images_4 (outdoor) and images_2 (indoor)
• OMMO , scene 01 resized to 1600px width
• Tanks & Temples, Truck and Train at their original resolution, without rescaling
• Deep Blending, Dr Johnson and Playroom

Note that the tandt_db.zip distributed with 3DGS is a different copy of the last two. Its Tanks & Temples images are downscaled, and its Deep Blending scenes are each missing one image, which shifts the test split. Results obtained from it will not match the ones reported here.

Your own data

Check out the 3DGS code for the preprocessing needed.

🎓 Citation

@inproceedings{baranowski2026conegs, 
    title={ConeGS: Error-Guided Densification Using Pixel Cones for Improved Reconstruction with Fewer Primitives}, 
    author={Bartłomiej Baranowski and Stefano Esposito and Patricia Gschoßmann and Anpei Chen and Andreas Geiger},
    year={2026},
    booktitle = {2026 International Conference on 3D Vision (3DV)}, 
}