Precise Forecasting of Sky Images Using Spatial Warping

Leron Julian, Aswin Sankaranarayanan

Image Science Lab, Carnegie Mellon University

Paper  Dataset 

SkyNet imrpoves sky-image prediction to model cloud dynamics with higher spatial and temporal resolution than previous works. Our method handles distorted clouds near the horizon of the hemispherical mirror by patially warping the sky images during training to facilitate longer forecasting of cloud evolution.

# To download dataset for train and test data:
pip install gdown
gdown --folder --id 1BkWx0j6Kt5G8CEMzzREprMeoYfw0v4ge

#If you run into an inssue requesting permission, update gdown first, then re-run the above command:
pip install --upgrade --no-cache-dir gdown

Installation

# Installation using using anaconda package management 
conda env create -f environment.yml
conda activate SkyNet
pip install -r requirements.txt

# How to train the model with default parameters:
python train.py

# For info about command-line flags use
python train.py --help

# NOTE: Make sure that your Pytorch, CUDA, and CuPuy Version Match and install Python Versiion >= 3.8:
pip install cupy-cuda11x

Thanks

This project makes use of LiteFlowNet for optical-flow estimates:

• LiteFlowNet2 for lightweight optical-flow estimates using a CNN Please refer to their webpage for installation and implementation

Citation

If you use this project in your research please cite:

@INPROCEEDINGS{SkyNet:ICCVW21,
  author={Julian, Leron and Sankaranarayanan, Aswin C.},
  booktitle={2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW)}, 
  title={Precise Forecasting of Sky Images Using Spatial Warping}, 
  year={2021},
  }