The EUVP dataset

Name: The EUVP (Enhancing Underwater Visual Perception) datasaet
Data: Paired and unpaired underwater images of poor and good perceptual quality
Usage: Training deep models for underwater image enhancement

Details about the dataset can be found in this paper; please cite it as: 

@article{islam2019fast,
    title={Fast Underwater Image Enhancement for Improved Visual Perception},
    author={Islam, Md Jahidul and Xia, Youya and Sattar, Junaed},
    journal={arXiv preprint arXiv:1903.09766},
    year={2019}
}

Paired data:

The paired dataset is prepared using a procedure suggested in our earlier work by Fabbri et al. We have recently refined and updated this data collectively. Two paired datasets are included, the statistics are as follows:

Dataset Name Training Pairs Validation Total Images
Underwater Dark 7200 675 15,075
Underwater ImageNet 4260 1400 9,920
Underwater Scenes 2325 142 4,792

 

Unpaired data:

The unpaired data is collected during oceanic explorations and human-robot collaborative experiments that are performed in different locations under various visibility conditions. Seven different cameras, which include multiple GoPros, Aqua AUV’s uEye cameras, low-light USB cameras, and Trident ROV’s HD camera, have been used to collect these data. Additionally, images extracted from a few publicly available YouTube videos are included in the dataset as well. The data is then labeled (i.e., poor or good quality) based on visual inspection by six individuals. The statistics are as follows:

Poor quality Good quality Validation Total Images
6445 7445 420 14,310