Visual Synthetic Data Generation for Sign Language Recognition

One of the essential steps in pattern recognition is having a training set that contains diversity of exemplars. This step is a major obstacle in sign language recognition. The availability of large public corpus with diversity of examples is limited due to the large number of degrees of freedom inherent in sign language hand gestures and the major variances between different signers. This causes difficulties in collecting a large number of exemplars for each hand gesture and makes it either expensive or impractical. This paper attempts to address this challenge by using hand motion carried out by synthetic 3D animated models. The experiments examine how the recognition accuracy will be changed when the training set is enlarged. The results demonstrate how the recognition accuracy has been improved even when the expressing ability of the classifier is limited. Moreover, it enhances the quality of training set by including a signer independent manner