README.md

metadata

license: mit
pretty_name: P3M-10K
task_categories:
  - image-segmentation
tags:
  - image-matting
  - portrait-matting
  - alpha-matte
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: P3M_500_P
        path: data/P3M_500_P-*
      - split: P3M_500_NP
        path: data/P3M_500_NP-*

P3M-10K

P3M-10K (Privacy-Preserving Portrait Matting) is a large-scale portrait matting benchmark. It is redistributed here from the original release by JizhiziLi/P3M.

If you use this dataset, please cite the original paper:

Jizhizi Li, Sihan Ma, Xin Zhang, Dacheng Tao. "Privacy-Preserving Portrait Matting." ACM International Conference on Multimedia (ACM MM), 2021.

Contents

Each example is a portrait RGB image and its corresponding alpha matte:

Column	Type	Description
image	Image	RGB portrait image
mask	Image	Alpha matte (soft foreground opacity)

Note: mask is the alpha matte, not a binary segmentation map.

Splits

Split	Rows	Image source	Description
train	9421	blurred_image (privacy)	Training set with face-blurred portraits
P3M_500_P	500	blurred_image (privacy)	Validation set, Privacy (face-blurred) portraits
P3M_500_NP	500	original_image	Validation set, Non-Privacy (unblurred) portraits

The schema is identical across all three splits (image + mask).

Dropped columns

The original release also ships fg/ (foreground) and bg/ (background) for the training set, and trimap/ for the validation sets. These were dropped to keep a clean, consistent two-column schema (image, mask) across all splits.

License

Released under the MIT License, matching the original P3M-10K Dataset Release Agreement (MIT License). Please review and abide by the original agreement.