Clean-GS: Semantic Floater Removal for 3D Gaussian Splatting

Official implementation of Clean-GS, a method for removing background clutter and floaters from 3D Gaussian Splatting reconstructions using sparse semantic masks.

Overview

3D Gaussian Splatting produces high-quality scene reconstructions but generates hundreds of thousands of spurious Gaussians (floaters) scattered throughout the environment. Clean-GS addresses this by combining whitelist-based spatial filtering with color-guided validation and outlier removal to achieve 60-80% model compression while preserving object quality.

Features

• Sparse Supervision: Uses as few as 3 segmentation masks (1% of views)
• Three-Stage Pipeline:
  1. Whitelist filtering via projection to masked regions
  2. Depth-buffered color validation
  3. Neighbor-based outlier removal (k-NN)
• High Compression: 60-80% reduction in Gaussians with maintained quality
• Monument Isolation: Extract clean objects from complex outdoor scenes

Installation

# Clone the repository
git clone https://github.com/YOUR_USERNAME/clean-gs.git
cd clean-gs

# Install dependencies
pip install -r requirements.txt

Requirements

• Python 3.8+
• PyTorch
• Open3D
• NumPy
• SciPy
• PIL

See requirements.txt for complete dependencies.

Data Structure

Your data should be organized as follows:

data/
├── datasets/
│   └── temple/                    # Scene name
│       ├── images/                # Input images
│       ├── sparse/0/              # COLMAP sparse reconstruction
│       │   ├── cameras.bin
│       │   ├── images.bin
│       │   └── points3D.bin
│       ├── cameras.json           # Camera parameters (generated)
│       └── masks/                 # Semantic segmentation masks (PNG)
│           ├── 000000.png
│           ├── 000001.png
│           └── ...
└── models/
    └── temple/                    # Trained 3DGS model
        └── point_cloud/
            └── iteration_30000/
                └── point_cloud.ply

Creating Masks

Use Segment Anything (SAM) to create semantic masks for your objects of interest. You only need 3-5 masks for effective floater removal.

Usage

Basic Usage

python clean-gs.py \
    --scene temple \
    --masks_dir data/datasets/temple/masks \
    --input_ply data/models/temple/point_cloud/iteration_30000/point_cloud.ply \
    --output_ply data/models/temple/point_cloud/iteration_30000/clean.ply

With Custom Parameters

python clean-gs.py \
    --scene temple \
    --masks_dir data/datasets/temple/masks \
    --input_ply data/models/temple/point_cloud/iteration_30000/point_cloud.ply \
    --output_ply clean_output.ply \
    --color_threshold 0.3 \
    --k_neighbors 5 \
    --neighbor_threshold 0.7

Parameters

• --scene: Scene name (must match directory in data/datasets/)
• --masks_dir: Directory containing semantic masks
• --input_ply: Input 3DGS point cloud file
• --output_ply: Output cleaned point cloud file
• --color_threshold: Color validation threshold (default: 0.3)
• --k_neighbors: Number of neighbors for outlier detection (default: 5)
• --neighbor_threshold: Outlier removal threshold (default: 0.7)

Results

Temple (Tanks and Temples)

• Original: 526K Gaussians, 125 MB
• Clean-GS: 198K Gaussians, 47 MB (62% reduction)
• Quality: Maintained rendering quality with PSNR preserved

Isha Statue

• Original: 1,112K Gaussians, 263.6 MB
• Clean-GS: 223K Gaussians, 52.8 MB (80% reduction)

Pipeline Stages

Stage	Removed	Remaining	Compression
Original	-	526K	0%
+ Whitelist	163K	363K	31%
+ Color Validation	155K	208K	60%
+ Neighbor Removal	10K	198K	62%

Quick Start

1. Prepare Data

See data/README.md for detailed data preparation instructions. Quick summary:

1. Capture or download multi-view images (100-300 views)
2. Run COLMAP to get sparse reconstruction

   bash scripts/run_colmap_temple.sh

3. Train 3D Gaussian Splatting model

   python train.py -s data/datasets/YOUR_SCENE -m data/models/YOUR_SCENE

4. Generate 3-5 semantic masks using SAM
5. Convert camera parameters

   python scripts/colmap_to_json.py --colmap_path data/datasets/YOUR_SCENE/sparse/0 --output_json data/datasets/YOUR_SCENE/cameras.json

2. Run Clean-GS

python clean-gs.py \
    --scene YOUR_SCENE \
    --masks_dir data/datasets/YOUR_SCENE/masks \
    --input_ply data/models/YOUR_SCENE/point_cloud/iteration_30000/point_cloud.ply \
    --output_ply clean_output.ply

Citation

If you use Clean-GS in your research, please cite:

@article{cleangs2026,
  title={Clean-GS: Semantic Floater Removal for 3D Gaussian Splatting},
  author={Subhankar Mishra},
  journal={arXiv preprint},
  year={2026}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• 3D Gaussian Splatting for the base representation
• Segment Anything for mask generation
• Tanks and Temples for benchmark datasets

Contact

For questions or issues, please open an issue on GitHub.