Gaussian Splatting Pipeline with Georeferencing and Correct Scaling

This repository provides a pipeline for Gaussian Splatting with georeferencing and accurate scaling. Starting from drone images that contain GPS EXIF data, this workflow produces 3D reconstructions that are nearly perfectly scaled and aligned with real-world coordinates. The reference frame is anchored to the GPS position of the first image in the dataset.

The following images illustrate the results: on the left is the Gaussian Splatting reconstruction, and on the right is the SuGaR-refined Gaussian representation, both uploaded to Cesium Ion.

Technologies Used

• COLMAP – Performs Structure-from-Motion (SfM) reconstruction to generate georeferenced sparse 3D point clouds from drone images. These point clouds are georeferenced and scaled according to GPS data.
• Gaussian Splatting – A neural rendering method that converts 3D points into Gaussian representations, producing highly detailed and accurate 3D reconstructions.
• SuGaR – A framework built around Gaussian Splatting that orchestrates the full pipeline:
  • Short vanilla 3DGS optimization – Optimizes a vanilla 3D Gaussian Splatting model for 7k iterations to position Gaussians in the scene.
  • SuGaR optimization – Refines Gaussian positions and aligns them to the surface of the scene.
  • Mesh extraction – Extracts a mesh from the optimized Gaussians.
  • SuGaR refinement – Builds a hybrid representation combining Gaussians and mesh for maximum accuracy.
  • Textured mesh extraction (optional) – Produces a traditional textured mesh for visualization, composition, and animation in Blender.

Docker Installation

For reproducibility and ease of use, the pipeline is provided in a Docker image.

Before running the container, you need to organize your local workspace. This ensures your data and outputs persist even after you stop or rebuild the Docker image.

1. Create the Workspace Structure

On your host machine, create a main project folder. The structure should look like this:

/workspace/
 ├── images/                 # Folder containing your raw drone images (with GPS EXIF data)
 └── georeferenced_gsplat/   # Folder where the processing scripts and outputs will live

2. Clone the repository

Inside the /workspace folder, clone the repository:

git clone https://github.com/manudelu/georeferenced_gsplat.git

3. Build the Docker image

docker build -t georeferenced_gsplat:latest .

Important: Before building, make sure to update the Dockerfile:

• Set ENV TORCH_CUDA_ARCH_LIST to match your GPU(s) compute capability
• Set -DCMAKE_CUDA_ARCHITECTURES in any CMake commands to your GPU(s) compute capability

4. Run the Docker Container

docker run --gpus all -it --name sugar-env -v /workspace:/home/workspace gaussian-splatting:22 bash

Pipeline Workflow

Launch the Pipeline

Make the script executable and run it inside the container:

cd /home/workspace/georeferenced_gsplat/scripts
chmod +x pipeline.sh
./pipeline.sh

If running on a remote server (e.g., via SSH):

nohup ./pipeline.sh > pipeline.log 2>&1 &
tail -f pipeline.log  # To monitor progress

Automated Processing Stages

1. Data Preparation

• Copy raw drone images from /home/workspace/images into /home/workspace/data/input.
• Extract GPS EXIF data from images into a text file (geotags.txt) using exif_to_txt.py.

2. COLMAP Reconstruction

• Run COLMAP to perform Structure-from-Motion (SfM) reconstruction.
• Convert the reconstruction to an ENU (East-North-Up) coordinate frame using the GPS position of the first image as the reference.
• Produces a georeferenced and scaled sparse point cloud aligned with real-world coordinates.

3. Gaussian Splatting Training

• Trains the Gaussian Splatting model using the georeferenced point cloud.
• Saves the results in /home/workspace/georeferenced_gsplat/output/vanilla_gs

4. SuGaR Training and Refinement

• Runs the SuGaR pipeline, refining the Gaussian scene to produce high-fidelity results.
• Generates:
  • Refined Gaussian point clouds (.ply)
  • UV-textured meshes (.obj)

5. Output Export

• Copies all final outputs from /home/SuGaR/output to /home/workspace/georeferenced_gsplat/output.
• The results are ready for visualization in Cesium Ion, CloudCompare, Unreal Engine, or other 3D/GIS software.

Importing into Cesium Ion

1. Login to Cesium Ion.
2. Go to My Assets -> Add Data.
3. Upload the .ply generated by SuGaR (refined point cloud).
4. To georeference:
   • Use the GPS coordinates of the first image in the dataset.
   • Adjust location in Cesium Ion and save.

Your Gaussian Splatting reconstruction will now be correctly aligned to real-world coordinates.

Importing into Unreal Engine

1. Open your Unreal Engine project (with Cesium for Unreal and LumaAI or XScene plugin installed).
2. Import the refined .ply into Unreal Engine.
3. Drag it into the Viewport.
4. Add a Cesium Globe Anchor actor.
   • Make the point cloud a child of the globe anchor or other 3D tiles (e.g., Google Photorealistic 3D Tiles).
   • Set the location to the GPS coordinates of the first image.
   • Adjust rotation if necessary.

Now your Gaussian Splatting reconstruction is correctly georeferenced in Unreal Engine.

Notes

• The pipeline automatically detects available GPUs for SuGaR training. If you want to force CPU-only execution or specify a different GPU, you can edit the pipeline.sh file.
• Mounting the /workspace directory ensures your data, scripts, and outputs persist across Docker rebuilds, keeping your experiments reproducible, organized, and clean.