Naver Webtoon Datasets

Dataset Construction Pipeline

0. Face detector

   Train YOLOv5 on the combination of two datasets

    iCartoonFace: https://github.com/luxiangju-PersonAI/iCartoonFace
    AnimeFaces: https://github.com/qhgz2013/anime-face-detector
   

   for about 120 epochs (mAP 0.979)

   Detection pipeline

    → full webtoon image (usually height >> width)
    → ratio-preserving resize to width 640 
    → split images horizontally (height 640) with overlaps (overlap = minimum detection size / 2)
    → run detection on splitted images
    → remove duplicated detections in overlaps
   

1. Filtering

   Run detector and filter images with criteria

    score > 0.82
    short edge size > 64
    aspect ratio < 1.3
   

   Then,

    → make square
    → expand 1.8 times
    → push down 0.2 (to roughly match the scale & alignment of CelebA-HQ and FFHQ)
    → crop and filter out if black/white blanks on the edge is > edge * 0.075
   

   Finally, use CLIP to fileter out images with texts

    labels = ["a face illustration without texts", "a face illustration with texts"]
   

2. Post Processing (included in the code)

   Waifu2x to upscale & denoise → resize to 256

3. To Do

   • semantic segmentation & landmark detection
   • full-body / objects