- FROSI dataset (DOI: 10.1109/IVS.2014.6856535): Synthetic images for classification. Visibility values range from 50 m to 400 m (7 types). The original image size in FROSI is 1400x600. We re-sample the images into the size of 140x60.
- Download the datasets, and then resample images in the dataset to the size of 140x60.
- Randomly split the datasets to training, validation, and test sets.
- Example images:
Visibility (m) | 50 | 100 | 150 | 200 | 250 | 300 | 400 |
---|---|---|---|---|---|---|---|
~ | |||||||
~ | |||||||
~ |
NB The link for FROSI dataset is not working. You can obtain the dataset by contacting the authors or use the one we provide. The images in the dataset https://github.com/cvvsu/RMEP/releases/tag/FROSIv0.0 we provide are 10x smaller than the original ones. Make sure that you should cite the paper (not our paper) DOI: 10.1109/IVS.2014.6856535 if you use the dataset:)
- SSF dataset is contructed based on 10.1109/WACV.2016.7477637 and 10.1109/WACV.2018.00032. The SSF dataset contains 24,328 images.
- Download the dataset from this link (85.2 MB)
- All images in the SSF dataset are with the size of 160x120.
- Please also cite the above-mentioned two papers if you use the SSF dataset.
- Webcam locations
- The original datasets including the labels can be downloaded from this website (3.3GB)
- Example images:
- Example images in RGBHSI
train
$ python3 main.py --name FROSI --dataroot datasets/FROSI --preprocess crop_and_flip --crop_size 60 --input_nc 3 --model_type clf_multi --output_nc 7
test
$ python3 main.py --name FROSI --dataroot datasets/FROSI --input_nc 3 --model_type clf_multi --output_nc 7 --isTest
train
$ python3 main.py --name RGB_128 --dataroot datasets/RGB_128 --output_params visibility RH temperature PM25 PM10 --preprocess crop_and_flip --crop_size 126 --input_nc 3 --model_type reg
test
$ python3 main.py --name RGB_128 --dataroot datasets/RGB_128 --input_nc 3 --model_type reg --output_params visibility RH temperature PM25 PM10 --isTest
Compared with other models on RGB_128 dataset:
Our code is modified from the PyTorch CycleGan code. Please also check their licence file.
The code is released under the Apache 2.0 license. Please refer to the LICENSE file for more information.