RMFAT [AAAI 2026]

Recurrent Multi-scale Feature Atmospheric Turbulence Mitigator

Zhiming Liu, Nantheera Anantrasirichai

📑 Contents

• Environment Installation
• My Dataset
• Training
• Performance Evaluation

🔨 Environment Installation

conda create -n RMFAT python=3.11
conda activate RMFAT
cd code
pip install -r requirements.txt

🧩 My Datasets

This is a public dataset for turbulence migitation tasks, containing two subsets: static and dynamic.

📦 Download

You can download each dataset subset as a ZIP file below:

• 👉 Download Static Dataset
• 👉 Download Dynamic Dataset

🛠️ Training

For the training on dynamic scene data, run the following:

python recursive_train.py --train_path "/path/to/dynamic/train/data" --val_path "/path/to/dynamic/val/data" --batch_size 1 --patch_size 256 --num_frames 10 --tmt_dims 16 --log_path "/path/to/save/logs/dynamic" run_name "train_dynamic" --resume_ckpt "/path/to/dynamic/checkpoint.pth"

For the training on static scene data, run the following:

python recursive_train.py --train_path "/path/to/static/train/data" --val_path "/path/to/static/val/data" --batch_size 1 --patch_size 256 --num_frames 10 --tmt_dims 16 --log_path "/path/to/save/logs/dynamic" --run_name "train_dynamic" --resume_ckpt "/path/to/dynamic/checkpoint.pth"

🚀 Performance Evaluation

For the single video inference, run the following:

python restore_video.py --video_path /path/to/video.mp4 --output_dir /path/to/output --restoration_ckpt /path/to/model.pth --tmt_dim 32 --warp_mode enc --n_frames 3 --save_raw

Citation

@inproceedings{liu2026rmfat,
  title     = {{RMFAT}: Recurrent Multi-scale Feature Atmospheric Turbulence Mitigator},
  author    = {Liu, Zhiming and Anantrasirichai, Nantheera},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  year      = {2026},
}