CelebA

Overview

CelebFaces Attributes Dataset (CelebA) is a large-scale face attributes dataset with more than 200K celebrity images, each with 40 attribute annotations. The images in this dataset cover large pose variations and background clutter. CelebA has large diversities, large quantities, and rich annotations, including

• 10,177 number of identities
• 202,599 number of face images, and
• 5 landmark locations, 40 binary attributes annotations per image.

The dataset can be employed as the training and test sets for the following computer vision tasks: face attribute recognition, face recognition, face detection, landmark (or facial part) localization, and face editing & synthesis.

Citation

@inproceedings{liu2015faceattributes,
  title = {Deep Learning Face Attributes in the Wild},
  author = {Liu, Ziwei and Luo, Ping and Wang, Xiaogang and Tang, Xiaoou},
  booktitle = {Proceedings of International Conference on Computer Vision (ICCV)},
  month = {December},
  year = {2015} 
}

Contact

Please contact Ziwei Liu and Ping Luo for questions about the dataset.

---

By Multimedia Lab, The Chinese University of Hong Kong

For more information about the dataset, visit the project website:

http://personal.ie.cuhk.edu.hk/~lz013/projects/CelebA.html

If you use the dataset in a publication, please cite the paper below:

@inproceedings{liu2015faceattributes, author = {Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang}, title = {Deep Learning Face Attributes in the Wild}, booktitle = {Proceedings of International Conference on Computer Vision (ICCV)}, month = December, year = {2015} }

Please note that we do not own the copyrights to these images. Their use is RESTRICTED to non-commercial research and educational purposes.

======================== Change Log

Version 1.0, released on 28/09/2015 Version 1.1, released on 23/03/2016, add landmarks annotations for align&cropped images Version 1.2, released on 08/04/2016, add align&cropped images in lossless format Version 1.3, released on 29/07/2016, add bounding box annotations for in-the-wild images Version 2.0, released on 28/06/2017, add identity annotations note

======================== File Information

• In-The-Wild Images (Img/img_celeba.7z) 202,599 original web face images. See In-The-Wild Images section below for more info.

• Align&Cropped Images (Img/img_align_celeba.zip & Img/img_align_celeba_png.7z) 202,599 align&cropped face images. See Align&Cropped Images section below for more info.

• Bounding Box Annotations (Anno/list_bbox_celeba.txt) bounding box labels. See BBOX LABELS section below for more info.

• Landmarks Annotations (Anno/list_landmarks_celeba.txt & Anno/list_landmarks_align_celeba.txt) 5 landmark location labels. See LANDMARK LABELS section below for more info.

• Attributes Annotations (Anno/list_attr_celeba.txt) 40 binary attribute labels. See ATTRIBUTE LABELS section below for more info.

• Identity Annotations (available upon request) 10,177 identity labels. See IDENTITY LABELS section below for more info.

• Evaluation Partitions (Eval/list_eval_partition.txt) image ids for training, validation and testing set respectively. See EVALUATION PARTITIONS section below for more info.

========================= In-The-Wild Images

------------ img_celeba.7z ------------

folder: img_celeba.7z.001, img_celeba.7z.002, ..., img_celeba.7z.014

---

Notes:

1. Please unzip these files together.

---

========================= Align&Cropped Images

------------ img_align_celeba.zip ------------

format: JPG

------------ img_align_celeba_png.7z ------------

format: PNG folder: img_align_celeba_png.7z.001, img_align_celeba_png.7z.002, ..., img_align_celeba_png.7z.016

---

Notes:

1. Images are first roughly aligned using similarity transformation according to the two eye locations;
2. Images are then resized to 218*178;
3. Please unzip "img_align_celeba_png.7z.*" together.

---

========================= BBOX LABELS

------------ list_bbox_celeba.txt ------------

First Row: number of images Second Row: entry names

Rest of the Rows: <image_id> <bbox_locations>

---

Notes:

1. The order of bbox labels accords with the order of entry names;
2. In bbox location, "x_1" and "y_1" represent the upper left point coordinate of bounding box, "width" and "height" represent the width and height of bounding box. Bounding box locations are listed in the order of [x_1, y_1, width, height].

---

========================= LANDMARK LABELS

------------ list_landmarks_celeba.txt ------------

First Row: number of images Second Row: landmark names

Rest of the Rows: <image_id> <landmark_locations>

------------ list_landmarks_align_celeba.txt ------------

First Row: number of images Second Row: landmark names

Rest of the Rows: <image_id> <landmark_locations>

---

Notes:

1. The order of landmark locations accords with the order of landmark names;
2. The landmark locations in "list_landmarks_celeba.txt" are based on the coordinates of in-the-wild images;
3. The landmark locations in "list_landmarks_align_celeba.txt" are based on the coordinates of align&cropped images.

---

========================= ATTRIBUTE LABELS

--------------- list_attr_celeba.txt --------------

First Row: number of images Second Row: attribute names

Rest of the Rows: <image_id> <attribute_labels>

---

Notes:

1. The order of attribute labels accords with the order of attribute names;
2. In attribute labels, "1" represents positive while "-1" represents negative.

---

========================= IDENTITY LABELS

---

Notes:

1. The face identities are released upon request for research purposes only. Please contact us for details;
2. There are no identity overlapping between CelebA dataset and LFW dataset.

---

========================= EVALUATION PARTITIONS

------------- list_eval_partition.txt -------------

All Rows: <image_id> <evaluation_status>

---

Notes:

1. In evaluation status, "0" represents training image, "1" represents validation image, "2" represents testing image;
2. Identities of face images are NOT overlapped within this dataset partition;
3. In our ICCV 2015 paper, "LNets+ANet" is trained with in-the-wild images, while "FaceTracer" and "PANDA-l" are trained with align&cropped images;
4. Please refer to the paper "Deep Learning Face Attributes in the Wild" for more details.

---