SkeFormer: Sign Language Segmentation Based on Skeleton Input Transformer

The work aimed to segment each gesture in the sign language video dataset collected by our school. To effectively capture the fine-grained features of gesture changes, a segmentation model based on skeleton input is adopted. We propose a sign language segmentation model combining ST-GCN with ASFormer. Our framework was compared to two strong baselines(ST-GCN, MS-GCN) for three tasks, and the experimental results were all the better than the baseline. At the same time, the F1 score in the experimental results has been significantly improved, which means that the number of segmentation errors can be effectively reduced.