ASLAD-190K: Arabic Sign Language Alphabet Dataset

The ASLAD-190K dataset is an extensive collection containing 190,000 meticulously labeled RGB images representing 32 alphabets of the ArSL. To capture these images, we enlisted the help of two signers and utilized two different computer webcams, namely the HP HD camera and HP Truevision HD. The MediaPipe library was crucial in capturing RGB photos of various sizes.

During the data collection process, we took great care to introduce diversity into the images. This involved varying lighting conditions, distances (zoom in/out), timings (day and night), 2D and 3D image rotations, and backgrounds. As a result, the dataset now offers a significantly augmented image corpus that closely resembles real-world scenarios, encompassing increasingly intricate situations. This augmentation enhances the performance of classification models during training.

Data Description

The ASLAD-190K dataset (ASLAD-190K.zip) is a rich compilation of 190,000 meticulously annotated RGB images, encompassing 32 distinct classes (organized into 32 folders) representing the Arabic Sign Language alphabet. These images, stored in JPEG format, vary in resolution, with each class containing approximately 5,108 to 7,092 images, leading to an uneven distribution across the classes. Image collection involved two signers using two different webcams: the HP HD camera and the HP TrueVision HD. The CVZone library, along with OpenCV and MediaPipe, facilitated the capture of RGB images in diverse sizes. During the data collection phase, we deliberately altered lighting conditions (low, medium, high), distances (close, far), 2D and 3D rotations, and backgrounds (simple, complex). This comprehensive approach resulted in a significantly enhanced dataset that accurately reflects real-world conditions and incorporates increasingly complex scenarios, thereby improving the training of classification models. Figure 1 presents a selection of images illustrating various alphabets from our extensive ASLAD-190K dataset.

On another hand, by employing the Random Under-Sampling technique, we effectively addressed the dataset imbalance, resulting in a balanced collection of 3,000 images per class and a total of 96,000 images (32 classes × 3,000 images) (ASLAD-96K/ASLAD-96K.zip).

Furthermore, this balanced dataset of 96,000 images was used to extract additional features, including Hand Landmark Coordinates, which were identified using a keypoint detection model such as MediaPipe. We also extracted 26 geometric angles that capture the relationships between hand joints, providing vital geometric features for gesture classification. The newly extracted features have been saved in a CSV file, which is included with the collected dataset.

The CSV file (ASLAD-96K/ASLAD-96K_63Hand landmarks+26Angles+Label+ImgName.csv) contains data related to hand landmarks and extracted features used for gesture recognition. It consists of the following columns:

Hand Landmarks Coordinates (1:63): These columns represent the 3D coordinates (x, y, z) of 21 key points on the hand. Each key point has an associated (x, y, z) value, forming a total of 63 columns. These landmarks are extracted from the hand image using a keypoint detection model like MediaPipe.
Extracted 26 Angles: This set of columns represents 26 angles calculated between specific points on the hand. These angles capture the relationships between different hand joints, providing critical geometric features for gesture classification.
Label Number: This column contains the numeric label corresponding to the gesture or sign represented by the hand in the image ranges from 0 to 31. The labels are integers, with each unique number representing a different class of hand gestures.
Image Name: This column holds the file name of the image from which the hand landmarks were extracted. Each image is represented by its file name in the format image_name.jpg, linking the data to the corresponding visual representation.

📂 Dataset Files

This dataset contains two sub-datasets:

• ASLAD-190K
  File: ASLAD-190K.zip
  Description: 190,000 images of Arabic Sign Language alphabet.

• ASLAD-96K
  Files:

  • ASLAD-96K.zip
  • ASLAD-96K_63HandLandmarks+26Angles+Label+ImgName.csv
    Description: Balanced version with Hand Landmark Coordinates and Geometric Angles (96,000 images).
    The accompanying CSV file contains 63 Hand Landmarks + 26 Angles + Label + Image Name for each sample.

📑 Citation

If you use this dataset in your research, please ensure to cite the following original works:

-Boulesnane, A., Bellil, L. & Ghiri, M.G. A hybrid CNN-random forest model with landmark angles for real-time Arabic sign language recognition. Neural Comput & Applic 37, 2641–2662 (2025). https://doi.org/10.1007/s00521-024-10729-7

-Boulesnane, Abdennour; Bellil, Lyna; GHIRI, Maissoun Ghouzlen (2024), “ASLAD-190K: Arabic Sign Language Alphabet Dataset consisting of 190,000 Images”, Mendeley Data, V2, http://dx.doi.org/10.17632/2fgpn5dwgc.2

@article{Boulesnane2024,
  author = {Boulesnane,  Abdennour and Bellil,  Lyna and Ghiri,  Maissoun Ghouzlen},
  title = {A hybrid CNN-random forest model with landmark angles for real-time Arabic sign language recognition},
  volume = {37},
  ISSN = {1433-3058},
  DOI = {10.1007/s00521-024-10729-7},
  number = {4},
  journal = {Neural Computing and Applications},
  publisher = {Springer Science and Business Media LLC},
  year = {2024},
  month = dec,
  pages = {2641–2662}
}

@article{Boulesnane2024_ASLAD190K,
  author       = {Boulesnane, Abdennour and Bellil, Lyna and Ghiri, Maissoun Ghouzlen},
  title        = {ASLAD-190K: Arabic Sign Language Alphabet Dataset consisting of 190,000 Images},
  year         = {2024},
  publisher    = {Mendeley Data},
  version      = {V2},
  doi          = {10.17632/2fgpn5dwgc.2}
}