Access to the SITR Hijab Dataset

This dataset contains photos of identifiable people, including women without a hijab. Images like these are privacy-sensitive, and protecting them is exactly what this research is about. The data is released under CC BY 4.0. Please tell us who you are and how you plan to use it.

We ask that you: (1) do not re-host the images or show individual images or faces in publications; (2) do not use the data to identify, track or profile individuals, or to build surveillance or dress-code enforcement systems; (3) remove any image whose subject asks for it to be removed.

Log in or Sign Up to review the conditions and access this dataset content.

SITR Hijab Dataset (cleaned)

A cleaned and relabelled version of the Roboflow Universe Hijab Dataset by Team S (v4). It was used to train and evaluate the hijab classifier in SITR:

M. Obaid, H. Bkhaitan, D. Maali, S. Hammad, T. Thaher. An Identity-Aware Privacy-Preserving Deep Learning Framework for Culturally Sensitive Image Sharing. IJACSA, 17(4), 909–919, 2026.

Contents

Split Images hijab (0) no_hijab (1)
train 1,248 588 660
validation 63 36 27
test 880 612 268
total 2,191 1,236 955

Each row contains:

  • image: the JPEG image (Roboflow export, 640Γ—640).
  • label: image-level label used in the paper, 0 = hijab, 1 = no_hijab. This is the class of the first annotated person in the image.
  • objects: all annotated people in the image. bbox is in COCO format [x, y, width, height] (pixels), and each box has its own category. 184 images contain more than one person, and 32 of them mix both classes.
  • source_image: the original filename before Roboflow's export suffix. Rows with the same source_image are copies or augmentations of the same photo. They never cross splits (no photo appears in more than one split), so evaluation is not leaked. There are 1,365 unique source photos.
from datasets import load_dataset
ds = load_dataset("hadeelbkh/sitr-hijab-dataset")   # requires requesting access and `hf auth login`
print(ds["train"][0]["label"], ds["train"].features["label"].names)

How it was built

The original Roboflow export (3,082 images, Pascal VOC XML) was reviewed by hand:

  • Removed 891 images that had no annotated object, had inconsistent annotations, or were of poor quality.
  • Corrected mislabelled annotations, for example hijab-wearing people labelled as "no hijab".
  • Kept the remaining 2,191 images in the existing train / validation / test splits.

No further augmentation was applied for training. Images were only resized to 224Γ—224 and ImageNet-normalised.

Notes and limitations

  • Class balance differs by split. The test set is 70% hijab / 30% no-hijab, so report per-class or macro metrics, not accuracy alone.
  • The validation set is tiny (63 images). It is a weak basis for model selection.
  • Image-level labels are ambiguous for multi-person images. Use objects if you need per-person labels.
  • Geographic, age, skin-tone and hijab-style coverage was not audited. Other head coverings (caps, hats, medical coverings, non-Islamic headscarves) are rare or absent.

Ethical considerations

The images show real, identifiable people, and half of them show women without a hijab. That is the very kind of exposure SITR is designed to prevent, which is why access is gated. We ask users not to re-host the images, show individual images in publications, or use the data for identification, surveillance or enforcing dress codes. If you appear in this dataset and want an image removed, open a discussion on this repository and it will be removed.

License and attribution

Released under CC BY 4.0, the same license as the source dataset.

Citation

@article{obaid2026sitr,
  title   = {An Identity-Aware Privacy-Preserving Deep Learning Framework for Culturally Sensitive Image Sharing},
  author  = {Obaid, Mahmoud and Bkhaitan, Hadeel and Maali, Duha and Hammad, Saja and Thaher, Thaer},
  journal = {International Journal of Advanced Computer Science and Applications},
  volume  = {17},
  number  = {4},
  pages   = {909--919},
  year    = {2026},
  doi     = {10.14569/IJACSA.2026.0170488}
}
Downloads last month
96

Models trained or fine-tuned on hadeelbkh/sitr-hijab-dataset