Preliminary Study on Data Augmentation Methods for Sign Language Recognition

Sign language dataset is composed of video and corresponding sentence pairs, which reflect the characteristics of a visual language. Since constructing a sufficient amount of sign language dataset is challenging due to the high cost of collecting and creating videos, it becomes difficult to apply large language model approaches to the sign language translation tasks, unlike other natural language processing problems. In this paper, we suggest augmentation methods for sign language data to address the data scarcity issue. To categorize the methods in the context of the sign language translation task, we introduce three criteria: signer's individuality, filming environment, and grammar. Then, we analyze how the augmentation methods can be utilized in existing sign language translation datasets.