This code repository is for our paper "See the Wind: Wind Scale Estimation with Optical Flow and VisualWind Dataset"
Authors: Qin Zhang, Jialang Xu, Matthew Crane, Chunbo Luo
The purpose is to estimate the wind scale from videos. To address this problem, we build a visual dataset named VisualWind which is the first of its kind video dataset, collecting videos with trees swaying under various scales of wind from social media, public cameras and self-recording. The link for the dataset is as blow. https://sme.uds.exeter.ac.uk/folders/48caf5102d6196b9645fab1f46e494ec. An access key is needed to get to the dataset due to the server security policy. Please contact the author to get the access key if you are interested.
We propose a dual-branch deep learning model to estimate the wind scales in an end-to-end manner, consisting of a motion branch to extract motion features by optical flow, and a visual branch to extract visual features by convolutional operation, and achieving 86.69% accuracy on the proposed dataset.
