Sign Language Recognition—Dataset Cleaning for Robust Word Classification in a Landmark-Based Approach

Abstract

Communication barriers between hard-of-hearing and hearing individuals can be mitigated through advancements in sign language recognition (SLR) systems. These SLR systems can also improve the user experience of hard-of-hearing people when interacting with conversational systems that could emerge in the near future. This work explores a landmark-based approach for word classification within an SLR system. The study investigates the impact of a novel data-cleaning methodology on model performance during training. Specifically, a data cleaning process focused on video trimming and sign placement correction is shown to significantly improve dataset quality, resulting in more accurate classification. This cleaner data not only facilitated a more stable training process for the RNN model but also effectively delayed the onset of overfitting compared to a model trained on the original data. The findings highlight the critical role of data quality, particularly when dealing with the limitations inherent to small datasets commonly encountered in SLR tasks. The contribution of this study lies in demonstrating how targeted data cleaning enhances model stability and performance in resource-limited SLR systems.