WIDER Face#

Format specification#

WIDER Face dataset is a face detection benchmark dataset, that available for download here.

Supported types of annotation:

  • Bbox

  • Label

Supported attributes for bboxes:

  • blur:

    • 0 face without blur;

    • 1 face with normal blur;

    • 2 face with heavy blur.

  • expression:

    • 0 face with typical expression;

    • 1 face with exaggerate expression.

  • illumination:

    • 0 image contains normal illumination;

    • 1 image contains extreme illumination.

  • pose:

    • 0 pose is typical;

    • 1 pose is atypical.

  • invalid:

    • 0 image is valid;

    • 1 image is invalid.

  • occluded:

    • 0 face without occlusion;

    • 1 face with partial occlusion;

    • 2 face with heavy occlusion.

Import WIDER Face dataset#

Importing of WIDER Face dataset into the Datumaro project:

datum project create
datum project import -f wider_face <path_to_wider_face>

Directory with WIDER Face dataset should has the following structure:

<path_to_wider_face>
├── labels.txt  # optional file with list of classes
├── wider_face_split # directory with description of bboxes for each image
│   ├── wider_face_subset1_bbx_gt.txt
│   ├── wider_face_subset2_bbx_gt.txt
│   ├── ...
├── WIDER_subset1 # instead of 'subset1' you can use any other subset name
│   └── images
│       ├── 0--label_0 # instead of 'label_<n>' you can use any other class name
│       │   ├──  0_label_0_image_01.jpg
│       │   ├──  0_label_0_image_02.jpg
│       │   ├──  ...
│       ├── 1--label_1
│       │   ├──  1_label_1_image_01.jpg
│       │   ├──  1_label_1_image_02.jpg
│       │   ├──  ...
│       ├── ...
├── WIDER_subset2
│  └── images
│      ├── ...
├── ...

Check README file of the original WIDER Face dataset to get more information about structure of .txt annotation files. Also example of WIDER Face dataset available in our test assets.

Export WIDER Face dataset#

With Datumaro you can convert WIDER Face dataset into any other format Datumaro supports. Pay attention that this format should also support Label and/or Bbox annotation types.

Few ways to export WIDER Face dataset using CLI:

# Using `convert` command
datum convert -if wider_face -i <path_to_wider_face> \
    -f voc -o <output_dir> -- --save-media

# Through the Datumaro project
datum project create
datum project import -f wider_face <path_to_wider_face>
datum project export -f voc -o <output_dir> -- -save-media

Export WIDER Face dataset using Python API:

import datumaro as dm

dataset = dm.Dataset.import_from('<path_to_wider_face', 'wider_face')

# Here you can perform some transformation using dataset.transform or
# dataset.filter

dataset.export('output_dir', 'open_images', save_media=True)

Note: some formats have extra export options. For particular format see the docs to get information about it.

Export to WIDER Face dataset#

Using Datumaro you can convert your dataset into the WIDER Face format, but for succseful exporting your dataset should contain Label and/or Bbox.

Here example of exporting VOC dataset (object detection task) into the WIDER Face format:

datum project create
datum project import -f voc_detection <path_to_voc>
datum project export -f wider_face -o <output_dir> -- --save-media --image-ext='.png'

Available extra export options for WIDER Face dataset format:

  • --save-media allow to export dataset with saving media files (by default False)

  • --image-ext IMAGE_EXT allow to specify image extension for exporting dataset (by default - keep original)