Installation

Installing COLMAP

Setting up the CONDA environment

''' conda create -n splatting python=3.11 conda activate splatting pip install torch torchvision pip install open3d trimesh scikit-image opencv-python plyfile tqdm pip install gsplat imageio tyro viser pyyaml opencv-python pycolmap torchmetrics tensorboard scikit-learn matplotlib nerfview splines pip install git+https://github.com/rahul-goel/fused-ssim/ --no-build-isolation pip install imageio-ffmpeg pip install sam2 segment-anything '''

Viewing clustered Gaussian splats

After training and clustering, a clustered results file is written to <result_dir>/clustering/clustering_results.pkl and a clustered checkpoint can be produced from your checkpoint. You can visualize clustered splats in two ways:

• From the pipeline (master.py) — the script will attempt to locate the latest checkpoint under <result_dir>/ckpts and automatically launch an interactive viewer after clustering completes.

• Manually using the viewer helper:

python src/view_clustered_splats.py /path/to/ckpt_29999.pt /path/to/dense/clustering/clustering_results.pkl --viewer

This will save a clustered checkpoint (*_clustered.pt) next to the original checkpoint, render preview images into a cluster_previews folder, and (with --viewer) launch the GsplatViewer at http://localhost:8080.

If you prefer only previews without the web viewer, omit --viewer.

Note: the interactive viewer requires viser and other rendering dependencies available in your environment.

Issues

• Need to find a way to safely install COLMAP. Currently on some machine the automated call to colmap works, in other tosses errors realted to lacking solvers ('Can't use SPARSE_SCHUR sparse_linear_algebra_library_type = SUITE_SPARSE, because support was not enabled when Ceres Solver was built.') or lack of a channel for rendering the visual.