Skip to content

Dataset.md

Latest commit

 

History

History
75 lines (53 loc) · 4.81 KB

Dataset.md

File metadata and controls

75 lines (53 loc) · 4.81 KB
Raw

README for Dataset Preparation

We used three different datasets for our work as benchmarks: ZTE、T-OLED and P-OLED.

ZTE dataset was proposed by DISCNet, and you can see this link for details.

T-OLED and P-OLED datasets were proposed by Microsoft, and used as benchmark in UDC 2020 challenge, you can see this link for details.

Overview

ZTE:

Path Files Format Description
dataset/UDC_syn_data Main folder
├─  data_scripts 4 .py Python scripts to process downloaded dataset.
├─  synthetic_data Synthetic dataset.
      └─  GT Reprojected and cropped HDR images at 800×800.
             ├─  train 2,016 .npy Training split.
             └─  test 360 .npy Test split.
├─  real_data Real dataset.
      ├─  input 30 .npy Input UDC images at 3264×2448, captured by ZTE phone.
      ├─  CCM_txt 30 .txt Color correction matrix for each real input image.
      └─  jpg_ZTE 30 .jpg Camera output after built-in ISP.
├─  PSF PSF-related files.
      ├─  kernel_code 9 .npy Kernel code generated by PCA, d=5.
      ├─  kernel_info_list .txt List to specify iamges with corresponding path of PSF. To be generated by a script.
      └─  ZTE_new 9 .npy PSFs in various angles, size 800×800.

T-OLED and P-OLED:

Path Files Format Description
dataset/UDC_real_data Main folder
├─  mat 8 .mat Packaged UDC images.
├─  Real_data Unpacked train set for training.
      └─  Toled_train Training split of T-OLED dataset at 1024×2048.
             ├─  LQ 240 .png Low quality UDC images.
             └─  HQ 240 .png Ground Truth.
    └─  Poled_train Training split of P-OLED dataset at 1024×2048.
             ├─  LQ 240 .png Low quality UDC images.
             └─  HQ 240 .png Ground Truth.

Dataset Preprocess

ZTE:

  1. Download and unzip datasets into current directory:

    python data_scripts/download_dataset.py --dataset=UDC

  2. Generating synthetic data by apply convolution with PSF for UDC simulation:

    python data_scripts/data_simulation.py --data_path=. --psf_type=ZTE_new

  3. Generating info list of images together with corresponding PSF:

    python data_scripts/generate_info_list.py --data_path=. --psf_type=ZTE_new --save_dir=./PSF/kernel_info_list

You can download the dataset and reference their official tutorial here for details.

T-OLED and P-OLED:

Download and unzip datasets into ./dataset.

You can download the dataset from this link.

After downloading the datasets, please modify the dataset path settings of the yaml files in .options accordingly, you can refer to the file structure we proposed in the Overview section.