【CVPR'2024🔥】Turb-Seg-Res: A Segment-then-Restore Pipeline for Dynamic Videos with Atmospheric Turbulence

Useful Links

Links	Description
	Official project page with detailed information
	Link to the GitHub repository
	Link to the CVPR 2024 paper
	Repository for simulating atmospheric turbulence effects
	Dataset used in the project

Setup and Run

git clone https://github.com/Riponcs/Turb-Seg-Res.git
cd Turb-Seg-Res
pip install -r requirements.txt
python Demo.py

Contributions

• High Focal Length Video Stabilization: Stabilizes videos captured by high focal length cameras, which are highly sensitive to vibrations.
• Turbulence Video Simulation: Introduces a novel tilt-and-blur video simulator based on simplex noise for generating plausible turbulence effects with temporal coherence.
• Unsupervised Motion Segmentation: Efficiently segments dynamic scenes affected by atmospheric turbulence, distinguishing between static and dynamic components.

Usage (Demo.py)

The main script for running the demo is Demo.py. It processes a set of input images, applies stabilization, and saves the output images.

python Demo.py

Configuration

The following configuration settings can be adjusted in Demo.py:

• doStabilize: Enable or disable image stabilization.
• ProcessNumberOfFrames: Number of frames to process from the input images.
• resizeFactor: Factor to resize images.
• MaxStb: Maximum allowed pixel shift for image stabilization.
• path: Path to input images.
• savePath: Path to save output images.

Additional Resources

• : A repository for simulating atmospheric turbulence effects on images using 3D simplex noise and Gaussian blur.

Citation

If you find this work useful, please cite our CVPR 2024 paper:

@article{saha2024turb,
    title     = {Turb-Seg-Res: A Segment-then-Restore Pipeline for Dynamic Videos with Atmospheric Turbulence},
    author    = {Saha, Ripon Kumar and Qin, Dehao and Li, Nianyi and Ye, Jinwei and Jayasuriya, Suren},
    booktitle = {Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
    year      = {2024},
}