Skip to content

Latest commit

 

History

History
70 lines (55 loc) · 3.29 KB

Dataset_Preparation.md

File metadata and controls

70 lines (55 loc) · 3.29 KB
Raw

Dataset Preparation

Resources

Here are some other resources to download the dataset:

Interface of Dataloader

We create general interfaces for FR and NR datasets in pyiqa/data/general_fr_dataset.py and pyiqa/data/general_nr_dataset.py. The main arguments are

  • opt contains all dataset options, including
    • dataroot_target: path of target image folder.
    • dataroot_ref [optional]: path of reference image folder.
    • meta_info_file: file containing meta information of images, including relative image paths, mos labels and other labels.
    • augment [optional] data augmentation transform list
      • hflip: flip input images or pairs
      • random_crop: int or tuple, random crop input images or pairs
    • split_file [optional]: train/val/test split file *.pkl. If not specified, will load the whole dataset.
    • split_index [optional]: default 1, which split to use, only valid when split_file is specified.
    • dmos max: some dataset use difference of mos. Set this to non-zero will change dmos to mos with mos = dmos_max - dmos.
    • phase: phase labels [train, val, test]

The above interface requires two files to provide the dataset information, i.e., the meta_info_file and split_file. The meta_info_file are .csv files, and has the following general format

- For NR datasets: name, mos(mean), std
    ```
    100.bmp   	32.56107532210109   	19.12472638223644
    ```

- For FR datasets: ref_name, dist_name, mos(mean), std
    ```
    I01.bmp        I01_01_1.bmp   5.51429        0.13013
    ```

The split_file are .pkl files which contains the train/val/test information with python dictionary in the following format:

{
    train_index: {
        train: [train_index_list]
        val: [val_index_list] # blank if no validation split
        test: [test_index_list] # blank if no test split
    }
}

The train_index starts from 1. And the sample indexes correspond to the row index of meta_info_file, starting from 0. We already generate the files for mainstream public datasets with scripts in folder ./scripts/.

Note that we generate train/val/test splits follow the principles below:

  • For datasets which has official splits, we follow their splits.
  • For official split which has no val part, e.g., AVA dataset, we random separate 5% from training data as validation.
  • For small datasets which requires n-split results, we use train:val=8:2 ratio.
  • All random seeds are set to 123 when needed.

Specific Datasets and Dataloader

Some of the supported datasets have different label formats and file organizations, and we create specific dataloader for them:

  • Live Challenge. The first 7 samples are usually removed in the related works.
  • AVA. Different label formats.
  • PieAPP. Different label formats.
  • BAPPS. Different label formats.

Test Dataloader

You may use tests/test_datasets.py to test whether a dataset can be correctly loaded.