README.md

metadata

title: 4dgs-dpm
app_file: app.py
sdk: gradio
sdk_version: 5.17.1

Sparse-view Spacetime Gaussian Splatting via Dynamic Point Maps

Installation

# Create environment
conda create -n ssgs python=3.10
conda activate ssgs

# Install PyTorch with CUDA
pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Usage

Go-Pro Multi-View Capture System

git clone the gp-control repo 

Web Interface (Recommended)

python app.py

Upload videos, adjust settings, and download results as ZIP.

Command Line

python vdpm/infer.py --input mv-video/your-videos --output output/vdpm
python -m gs.train --input output/vdpm --output output/splats --iterations 1000

Pipeline

1. Video Processing: Extract and interleave frames from multi-view videos
2. VDPM Inference: Generate dynamic point maps and camera poses using VGGT backbone
3. 3DGS Training: Train per-frame Gaussian splats initialized from point maps
4. Animation Rendering: Generate GIF from interpolated camera viewpoint

Output

The pipeline generates:

• splats/frame_XXXX.ply - Gaussian splat for each timestep
• renders/ - Training progress images
• animation.gif - Rendered animation from average camera
• tracks.npz - 3D point tracks
• poses.npz - Camera poses

Requirements

Tested on:

• Windows 11
• RTX 3070 Ti
• CUDA 11.8+
• Python 3.12

TO-DO

• VGGT Quantization (BF16/FP16)
• Co-visibility check to reduce points
• Dynamic point tracking
• Per-frame 3DGS training
• Gradio demo with GIF rendering
• Dynamic/Static segmentation
• 3DGS with dynamic deformation field
• 4DGS primitive support

Citation

@misc{dpmsplat2026,
  title={DPM-Splat: Video to 4D Gaussian Splats via Dynamic Point Maps},
  author={Your Name},
  year={2026},
  url={https://github.com/YOUR_USERNAME/4dgs-dpm}
}

Acknowledgements

• VGGT - Visual Geometry Grounded Transformer
• 3D Gaussian Splatting
• NVIDIA Warp