KSU-ArSL, Arabic Sign language

The dataset consists of 80 classes (belonging to 80 Arabic alphabets, numbers, and common words) recorded by 40 healthy subjects using three cameras (one RGB and two Microsoft Kinect cameras). Each subject repeated each sign 5 times in five separate sessions at the same day. As a result, there are 200 video samples per class, 16000 samples in total per camera. Dataset recording was performed in an uncontrolled environment to reflect real-world conditions. No restrictions were imposed on lighting conditions, background color, and people's clothing. The distances between the acquisition equipment and the subjects were variable. This restriction-free recording makes the KSU-ArSL dataset very challenging. KSU-ArSL dataset have been validated on different DL architectures, including 3D convolutional neural network (CNN), attention-based CNN, multi-branch CNN, skeleton-based CNN, and CNN models fused with others DL layers including long short-term memory (LSTM), autoencoder (AE), and multilayer perceptron (MLP).