Sign Language Database.zip

The dataset includes over 2000 images where 9 non-expert subjects executed 37 Urdu alphabet signs. and over 700 images where 9 non-expert subjects executed 12 Urdu words. Skeletal key points of each data point have been extracted in the form of JSON files by using OpenPose. JSON files are stored in a folder named as its label. main_dataset.db file contains normalized alphabet data in the rightHandDataset table and word data in the poseDataset table. rightHandDataset table has x,y coordinates of 21 right-hand key-points and a label. Data in this table is normalized by scaling and moving these key points to a fixed position while keeping the original shape of the hand. In order to minimize rotation and positional variance of signs, new data is synthesized by rotating the right hand to +20 and -20 degrees and storing them as two variations of the same sign. poseDataset table has x,y coordinates of 21 right hand, 21 left hand and 13 body key-points and a label. Data in this table is normalized by scaling the body and each hand, and by moving the body to a fixed position, left hand to left wrist point and right hand to right wrist point, while keeping the original shape of the body and the hands. In order to minimize rotation and positional variance of signs, new data is synthesized by moving the hands 80 pixels up, down, left and right while keeping the natural posture of arms using Inverse Kinematics(IK) and rotating the hands to +20 and -20 degrees and storing them as 12 different variations of the same sign. Maintainers