Anson-Saju-George/wms-rifev3

WMS-RIFE v3 — Fine-tuned for WMS Satellite Sequences

Paper (ECCV2022)

Introduction

This repo provides a fine-tuned version of RIFE v3 (HDv3) adapted for WMS satellite imagery (2k–5k). Our model is optimized for smooth temporal interpolation of clouds, land/sea surfaces, and atmospheric patterns.

• Base architecture: RIFE (ECCV 2022)
• Training improvements: L1 + motion-weighted Charbonnier + (1-SSIM) + flow smoothness
• Domain: remote sensing imagery (multi-sensor, 2k–5k resolution)
• Performance: runs in real-time on RTX-class GPUs

---

Results

Model	PSNR ↑	SSIM ↑
Baseline (HDv3)	32.57	0.8867
Fine-tuned (L1 only)	34.10	0.8974
Fine-tuned (Custom)	34.42	0.8991

N = ~5.3k validation triplets from WMS sensors. Visual inspection confirms improved temporal stability and sharper cloud edges.

---

Quick Start

Installation

git clone https://huggingface.co/Anson-Saju-George/wms-rifev3
cd wms-rifev3
pip install -r requirements.txt

Inference (pair of images)

python demo_infer_pair.py --img0 frame_000.png --img1 frame_001.png

This will output interp_000_001.png.

Inference (video)

python finetune_infer_wms_sequence.py --video sample.mp4 --exp 1

Generates sample_2X.mp4.

---

Training / Evaluation

• Train:

  python finetune_custom_loss_train.py
  

• Eval:

  python finetune_eval_wms.py
  

---

Applications

• Satellite nowcasting & forecasting previews
• Meteorology (cloud motion, precipitation fronts)
• Ocean/atmosphere flow visualization
• Geophysics time-lapse smoothing
• Web maps & GIS visualization overlays

---

Datasets

We trained on curated WMS imagery (2k–5k resolution). Due to license restrictions, raw datasets are not redistributed. Please build from your own WMS feeds.

---

Citation

If you use this model, please cite the base RIFE paper:

@inproceedings{huang2022rife,
  title={Real-Time Intermediate Flow Estimation for Video Frame Interpolation},
  author={Huang, Zhewei and Zhang, Tianyuan and Heng, Wen and Shi, Boxin and Zhou, Shuchang},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2022}
}

---

License

Apache-2.0 (same as original RIFE). Please also respect the license of any datasets you use.

---