Skip to content

Inria gaussian-splatting — How-To (Training, Export, and Visualization)

This document describes the workflow used for the Inria Gaussian Splatting implementation, including:

  • environment setup and installation (Windows)
  • dataset preparation from video
  • dataset conversion
  • training
  • visualization with SIBR (Windows)
  • export for downstream inspection and cleaning

References

For this work, the Windows-oriented fork was used as installation reference and the tutorial video was followed for the complete setup workflow.
Both resources are maintained by the same author.

1. Overview

The Inria gaussian-splatting implementation is the reference research implementation of 3D Gaussian Splatting.

The pipeline operates on calibrated camera poses.
If they are not provided, the convert.py script can be used to automatically run a Structure-from-Motion reconstruction using COLMAP.

The workflow is composed of two main stages:

  1. Dataset preparation (convert.py)
  2. Optimization (train.py)

The SIBR viewer is a separate real-time renderer used only for visualization.

2. Prerequisites

  • NVIDIA GPU with CUDA support
  • CUDA Toolkit compatible with your PyTorch build (tested with CUDA 11.8 and PyTorch 1.12.1)
  • Python 3.8+ (Conda/Miniconda recommended) (tested with Python 3.8.20)
  • Git

2.1 Windows-specific notes

The Windows fork and the tutorial video provide step-by-step guidance for:

  • installing the CUDA toolkit (CUDA 11.8)
  • ensuring CUDA + PyTorch compatibility
  • Visual Studio setup (VS2019 for compatibility with CUDA 11.8)
  • installing COLMAP
  • installing ImageMagick for image preparation
  • installing FFmpeg to extract images from videos
  • building CUDA extensions if pip install fails
  • downloading and using the SIBR pre-built binaries for Windows
  • common Windows build/runtime issues

3. Clone the Repository

Always clone the official repository with submodules:

git clone https://github.com/graphdeco-inria/gaussian-splatting.git --recursive
cd gaussian-splatting

4. Create the Environment

Using Conda:

SET DISTUTILS_USE_SDK=1  # Windows only
conda env create --file environment.yml
conda activate gaussian_splatting

5. Optional Dataset Preparation from Video (ffmpeg)

If the dataset is provided as a video rather than a set of images, frames must first be extracted.
This can be done using ffmpeg, as shown in the Windows tutorial video.

Run the following command from the directory containing the input video:

ffmpeg -i {VIDEO} -qscale:v 1 -qmin 1 -vf fps={FPS} %04d.jpg

Where:

  • {VIDEO} should be replaced with the filename or full path to the input video file
  • {FPS} should be replaced with the desired frame extraction rate

The extracted frames will be saved as sequentially numbered .jpg images (e.g., 0001.jpg, 0002.jpg, ...).

5.1 FPS Selection Heuristic

The extraction FPS was chosen so as to obtain approximately 150 frames per scene (151 frames in the final benchmark datasets).

6. Input Dataset

The Inria implementation supports two dataset preparation modes.

6.1 Precalibrated Dataset

The training pipeline can directly consume an existing SfM reconstruction.

Expected structure:

<location>
|---images
|---sparse
    |---0
        |---cameras.bin
        |---images.bin
        |---points3D.bin

In this mode, camera poses and intrinsics are already known and the Gaussian Splatting optimization runs directly on the calibrated dataset.

6.2 Automatic Reconstruction from Images

If only images are available, the convert.py script can automatically run COLMAP and prepare the dataset.

Place images inside:

<location>
|---input
    |---<image 0>
    |---<image 1>
    |---...

Then run:

python convert.py -s {DATA_PATH}

Example:

python convert.py -s data\video-interno

7. Training

From the repository root:

python train.py -s {DATA_PATH}

Example:

python train.py -s data\video-interno

Training outputs and checkpoints are written under the output/ directory.

8. Visualization with SIBR (Windows)

The Inria implementation provides a real-time OpenGL viewer called SIBR.

After downloading the pre-compiled Windows build as shown in the tutorial video, the SIBR viewer binaries were copied into the root directory of the gaussian-splatting repository.
The resulting executable path is:

<GAUSSIAN_SPLATTING_ROOT>\viewers\bin\SIBR_gaussianViewer_app.exe

To visualize a trained scene, move to the viewers\bin directory and run:

SIBR_gaussianViewer_app -m {DATA_PATH}

where -m points to the output directory generated by the training run.

Example:

SIBR_gaussianViewer_app -m output\video-interno

9. Export

The exported .ply file can be found in the output directory of the training run.