Pytorch-TestRankIQA

RankIQA was proposed in a ICCV2017 paper by Liu X. You can get this paper from arXiv or ICCV 2017 open access.

This repo contains RankIQA model files in Pytorch.
You can use it to test RankIQA on TID2013 and LIVE dataset in Pytorch.
If you just want to have a quick and simple comparison with RankIQA on your own test set, this repo is suited for you.

News

• We have released our recent work RecycleD. It has been accepted to ACM MM 2021 brave new ideas.
  It is an opinion-unaware non-reference IQA method which is based on the pre-trained discriminator of WGAN.
  It may interest you if you are trying to use or study RankIQA.

Prerequisites

• Win10 (Not tested on Ubuntu yet)
• Python 3.6
• Numpy 1.19.1
• Pytorch 1.2
• Opencv 4.5

The above versions are not mandatory, just because I ran the code in such an environment.

Getting Started

python main.py --test_set "ur_path/TID2013/" --model_file "./pre-trained/Rank_tid2013.caffemodel.pt" --test_file "./data/ft_tid2013_test.txt" --res_file "./result.csv"
python main.py --test_set "ur_path/TID2013/" --model_file "./pre-trained/FT_tid2013.caffemodel.pt" --test_file "./data/ft_tid2013_test.txt" --res_file "./result.csv"

python main.py --test_set "ur_path/LIVE2/" --model_file "./pre-trained/Rank_live.caffemodel.pt" --test_file "./data/ft_live_test.txt" --res_file "./result.csv"
python main.py --test_set "ur_path/LIVE2/" --model_file "./pre-trained/FT_live.caffemodel.pt" --test_file "./data/ft_live_test.txt" --res_file "./result.csv"

Note: test_set is the dataset folder, model_file is the pre-trained model file, test_file is the txt file which contains MOS and image filenames (see here), res_file is the csv file to save the test results.

About the pre-trained model files

I use caffemodel2pytorch to transform the Caffe model file to Pytorch format. You can find the pre-trained Caffe model files of RankIQA in here.

You can also download the Pytorch model files transformed by myself from:

• Baidu disk, and the password is riqa.
• Google drive.

I have used these pre-trained model files to test the performance on LIVE2 and TID2013.
The test results are recorded in test results.xlsx.

Tips

I cannot guarantee the correctness of the pre-trained Pytorch model files and the test results.

I just tried to reproduce the results showed in the paper, and you can see the reproduced results on TID2013 and LIVE.

If you are familiar with Pytorch, you can modify the code and test RankIQA on other datasets.
I didn't write the training code.