IQADataset

General dataset for Image Quality Assessment on PyTorch.

Introduction

This Python 3 library inherites torch.utils.data.Dataset .

Directly use it in cooperated with torch.utils.data.Dataloader.

Basic Usage

It is extremely easy to use this dataset class.

Just extract the dataset downloaded without any additional operation and pass the path to the class constructor.

For example:

from IQADataset import LIVE2016

if __name__ == '__main__':
    dataset = LIVE2016('/path/to/live2016')
    dataset[0] # -> (img, label, distoration_type)

img is torch.Tensor in "CHW" dimention order.

label is MOS or DMOS index.

The whole dataset is automaticly separated.

Training to Evaluating ratio is 8:2.

dataset.train() must be called before training and dataset.eval() before evaluating.

Different parts of dataset will be provided according to that.

Call dataset.all() to let the dataset provide all images.

Options

• If you prefer preloading the whole dataset into memory, specify using_data_pack=True.

• If you need reference image meanwhile, specify require_ref=True.

• You can specify crop_shape to apply random crop in train mode.

• Actually, the dataset implements 5-fold strategy. You can specify n_fold to reset it.

• Use dataset.set_i_fold(i) to specify which part is used for evaluting. Others are used for training.