PIPAL-dataset icon indicating copy to clipboard operation
PIPAL-dataset copied to clipboard

[ Official ] - PIPAL Dataset and Training Codebase. ECCV-2020, NTIRE-21/22.

Perceptual IQA Dataset (PIPAL) and Codebase

PIPAL: a Large-Scale Image Quality Assessment Dataset for Perceptual Image Restoration

Jinjin Gu, Haoming Cai, Haoyu Chen, Haoyu Chen, Jimmy S.Ren, Chao Dong. In ECCV, 2020.

🔥 Important Notes [ 2022-Jan-02 ]

  • [2022-Jan-02] Update the README. I will recorrect those errors mentioned in the issues recently. Moreover, we will extend the PIPAL dataset with more degradations for the NTIRE 2022 (We are still waiting for further notes from the NTIRE organizer).

  • [2021-Feb] We are organizing NTIRE 2021 Perceptual IQA Challenge !!.

  • [2021-Feb] ❗️ ❗️ This codebase ONLY supports users to train LPIPS on PIPAL or BAPPS for now. The SWD module will be added in the future.

📦 Download PIPAL NTIRE 2021

🔧 Dependencies and Installation

  • Python 3 (Recommend to use Anaconda)
  • PyTorch >= 1.0
  • NVIDIA GPU + CUDA
  • Python packages: pip install numpy opencv-python lmdb pyyaml
  • TensorBoard:
    • PyTorch >= 1.1: pip install tb-nightly future
    • PyTorch == 1.0: pip install tensorboardX

💻 How to Train

  • Train Your IQA
    1. Prepare IQA dataset PIPAL public training [NTIRE 2021] or BAPPS. More details are in codes/data.
    2. Modify the dataset format based on your need in codes/data/PairedTrain_dataset.py and ValidorTest_dataset.py
    3. Modify the configuration file codes/options/train_test_yml/train_our_IQA.yml
    4. Run command:
    ython train.py -opt options/train_test_yml/train_IQA.yml
    

📈 How to Test

  • Test your IQA
    1. Prepare IQA dataset PIPAL public validation [NTIRE 2021]. More details are in codes/data.
    2. Modify the dataset format based on your need in ValidorTest_dataset.py
    3. Modify the configuration file codes/options/train_test_yml/test_IQA.yml
    4. Run command:
    ython test.py -opt options/train_test_yml/test_IQA.yml
    

Ackowledgement

  • This code is based on mmsr.