PseudoDepth-SLR: Generating Depth Data for Sign Language Recognition

In this paper, we investigate the significance of depth data in Sign Language Recognition (SLR) and propose a novel approach for generating pseudo depth information from RGB data to boost performance and enable generalizability in scenarios where depth data is not available. For the depth generation, we rely on an approach that utilizes vision transformers as a backbone for depth prediction. We examine the effect of pseudo depth data on the performance of automatic SLR systems and conduct a comparative analysis between the generated pseudo depth data and actual depth data to evaluate their effectiveness and demonstrate the value of depth data in accurately recognizing sign language gestures. Our experiments show that our proposed generative depth architecture outperforms an RGB-only counterpart.