Аннотация
Sign Language Recognition (SLR) is a main technology for bridging the communication gap between the deaf community and the hearing majority. While deep learning has advanced SLR significantly, low resource languages like Kazakh Sign Language (KSL) remain under explored due to the deficit of labeled data. In this paper, we address this limitation by establishing a novel benchmark for KSL, focusing on two distinct tasks: Isolated Sign Language Recognition (ISLR) and Emotion Recognition. We evaluate the performance of three state-of-the-art Vision Transformer architectures ViViT, VideoMAE V2, and TimeSformer on a custom collected dataset comprising 20 lexical gestures and 4 emotional states. Our experiments reveal that TimeSformer achieves superior performance, attaining a Top-1 Accuracy of 96.63% on lexical gestures and 80.87% on emotion recognition. Comparative analysis indicates that TimeSformer's "Divided Space-Time Attention" mechanism captures finegrained spatiotemporal dynamics more effectively than the factorised encoder of ViViT or the masked modeling approach of VideoMAE.