camvis2 - Estimation of Visibility on Webcam Images using Multi-Magnification Convolutional Networks

This dataset contains the data files of the camera-based visibility estimation project proposed on the MeteoSwiss/camvis2 github repository.

The contents of this data repository are as follows:

1 (Necessary) Dataset creation files: These files are used to create the dataset on which the models are trained and evaluated. It is necessary to download the compressed archives containing the files, decompress them and merge them with the corresponding directory in the data/raw/ directory of the project repository. The dataset creation files are the following archives: depth_maps.tar.gz and images.tar.gz.

2 (Facultative) Preprocessed dataset files: If you want to avoid creating the dataset, you can also get the preprocessed version of the dataset files in the processed.tar.gz archive. Similary, If you want to use these files, you need to download the compressed archive, extract its content and put them in the data/processed directory of the project repository.

3 (Facultative) Model training outputs. These files might be helpful if you don't have access to a GPU or don't want to train the models yourself for another reason. In the checkpoints.tar.gz archive, you'll find the pretrained weights of our models for each of the three experiments we provide in the run_scripts/ directory of the project repository. The content of this archive belongs to outputs/checkpoints.

4 (Facultative) Model evalutation outputs. Mostly useful if you want to compare your results with ours. Again, you'll find the outputs that correspond to every experiment provided in run_scripts/, but this time at evaluation-time. The outputs are stored in the following archives: val_cf_matrices.tar.gz, val_histplots.tar.gz, val_images.tar.gz, val_scores.tar.gz. They are to be put into the corresponding subdirectories in the outputs/ directory, in the project repository.