README.md

---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: bbox
    list: int32
    length: 4
  - name: facial_landmarks
    struct:
    - name: lefteye
      list: int32
      length: 2
    - name: righteye
      list: int32
      length: 2
    - name: nose
      list: int32
      length: 2
    - name: leftmouth
      list: int32
      length: 2
    - name: rightmouth
      list: int32
      length: 2
  splits:
  - name: train
    num_examples: 45843
  - name: test
    num_examples: 5561
  - name: validation
    num_examples: 5442
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: validation
    path: data/validation-*
---

CelebA in-the-wild images, bounding boxes, and facial landmarks.

Images where the face covers less than 30% of the image area are filtered out, resulting in:

* Train size: 45843
* Validation size: 5442
* Test size: 5561

Original dataset: https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

Processed directly from the original CelebA in-the-wild image IDs, bounding boxes, landmarks, and image-size metadata so image, bbox, and landmark rows stay aligned by `image_id`.