CelebA dataset provides an aligned set img_align_celeba.zip. However, the size of each aligned image is 218x178, so the faces cropped from such images would be even smaller!

Here we provide a code to obtain higher resolution face images, by cropping the faces from the original unaligned images via 68 landmarks.

We also use a deep image quality assessment method to evaluate and rank the cropped image quality in scores.txt, lower score the better.

Notice: There are still some low resolution cropped faces since the corresponding original images are low resolution.

• Prerequisites

  • OpenCV

  • Python 3.6

• Dataset

  • CelebA-unaligned (10.2GB, higher quality than the aligned data)

    • download the dataset

      • img_celeba.7z (move to ./data/img_celeba.7z): Google Drive or Baidu Netdisk (password rp0s)

      • annotations.zip (move to ./data/annotations.zip): Google Drive

    • unzip the data

      7z x ./data/img_celeba.7z/img_celeba.7z.001 -o./data/
      
      unzip ./data/annotations.zip -d ./data/

• Cropping Examples

  • 512x512 + lanczos4 + jpg

    python align.py --crop_size_h 512 --crop_size_w 512 --order 4 --save_format jpg --n_worker 32

    

  • 512x512 + lanczos4 + png + larger face in the image (by setting face_factor, default is 0.45)

    python align.py --crop_size_h 512 --crop_size_w 512 --order 4 --save_format png --face_factor 0.6 --n_worker 32

    

  • 384x384 + bicubic + jpg + smaller face in the image (by setting face_factor, default is 0.45)

    python align.py --crop_size_h 384 --crop_size_w 384 --order 3 --save_format jpg --face_factor 0.3 --n_worker 32

    

• Notice

  • order

    • 0: INTER_NEAREST

    • 1: INTER_LINEAR

    • 2: INTER_AREA

    • 3: INTER_CUBIC

    • 4: INTER_LANCZOS4

    • 5: INTER_LANCZOS4