Compact3D (CompGS)

Training is performed using the run.sh script located in the gaussian-splatting directory.

Steps:

1. Open run.sh and modify the paths for your dataset and the output folder.
2. Execute the training script:

   cd gaussian-splatting
   bash run.sh

Optimization Tips: To reduce training time by 30%-50% at the cost of a minor performance reduction, you can:

• Decrease kmeans_iters (e.g., from 10 to 5).
• Increase kmeans_freq (e.g., to 200).

Opacity Regularization: By default, run.sh includes the opacity_reg argument. This regularizes opacity values and prunes highly transparent Gaussians, leading to further compression (up to 40-50x total) and faster rendering. To disable this, remove opacity_reg from the arguments list in run.sh.

Once training is complete, use the standard 3D Gaussian Splatting rendering and evaluation scripts.

Important: When working with quantized models, you must use the --load_quant flag. To evaluate an original non-quantized model, omit this flag.

Generate Renderings:

python render.py -m <path to trained model> --skip_train --load_quant

• --skip_train: Avoids rendering images from the training set camera viewpoints.
• --load_quant: Required to load the quantized model parameters.

Compute Metrics:

python metrics.py -m <path to trained model>

Standard 3D Gaussian Splat viewers cannot directly open quantized models. To visualize a trained model in an off-the-shelf viewer, you must first decompress it into a standard .ply file.

Run the following command to generate a file named point_cloud_decompressed.ply in the project's save path:

python decompress_to_ply.py -m <path to trained model>

CompGS is based on the official 3D Gaussian Splatting repository. To set up the environment, follow these steps:

1. Clone the CompGS repository:

   git clone https://github.com/UCDvision/compact3d
   cd compact3d

2. Install the official 3D Gaussian Splatting repository following its specific instructions.
3. Install the bitarray package to handle K-Means indices efficiently:

   pip install bitarray

4. After installing 3D Gaussian Splatting, ensure you have a gaussian-splatting directory inside your compact3d folder. Move the downloaded files from the CompGS repo to the appropriate locations within that folder using the provided script:

   bash move_files_to_gsplat.sh