Datasets preparation

We provide some examples about how to prepare the structure of data in this folder.

This is essential because the data structure is related to the data reader function and the loss function.

  1. Download the LIVE2 dataset and rename it as live folder. Rename the image name in each distortion folder from “imgx.bmp” and “imgxx.bmp” to “img0xx.bmp”, which allows us to read the data by order. 
    python rename_image_names.py
  2. Download the TID2013 dataset and rename it as tid2013 folder.
  3. Download the Waterloo or Validation set of Places2 datasets, and put it in rank_live and rank_tid2013 folders.
  4. Follow the instructions of Waterloo to generate 4 distortions (JPEG, JP2K, Gblur and GN) at five levels in rank_live. The code to generate 17 distortions for tid2013 dataset will be publicly available in the future.

Ranking datasets preparation

To train the RankIQA model on generated rank dataset, the data structure should be like tid2013_train.txt.

Run the code to generate the live_train.txt and live_test.txt files

cd data
python generate_rank_txt_live.py

Run the code to generate the tid2013_train.txt and tid2013_test.txt files

python generate_rank_txt_tid2013.py

IQA datasets preparation

To train the RankIQA+FT model finetuned on tid2013 dataset, the data structure should be like ft_tid2013_train.txt.

Run the code to generate the ft_live_train.txt and ft_live_test.txt files

python generate_ft_txt_live.py

Run the code to generate the ft_tid2013_train.txt and ft_tid2013_test.txt files

python generate_ft_txt_tid2013.py
cd ..