Scholarlink

This study compares three deep learning models for Indian Sign Language (ISL) interpretation, with a focus on how well they learn spatial and temporal patterns using video-based gesture data. To capture detailed motion dynamics, MediaPipe’s Holistic model extracted 543 landmark sequences from video frames, including 21 right and left hands, 468 face, and 33 posture landmarks. The models tested include an Attention-Based BiDirectional Long Short-Term Memory (Bi-LSTM) Network, a Pose-Guided Graph Convolutional Network (PGCN), and a Spatiotemporal Transformer for Vector Sequences (STraVe). The findings show that the Attention-Based Bi-LSTM Network had the greatest classification accuracy of 95.87%, outperforming PGCN (94.70%) and STraVe (87.16%), capturing the complex spatiotemporal aspects of ISL.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Comparative Analysis of Spatiotemporal Deep Learning Models for Indian Sign Language Recognition Using MediaPipe Holistic Landmarks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chayanika Basak,
  • Kanushree Anand,
  • Pratibha,
  • Pooja Kumari,
  • Shailesh D. Kamble

摘要

This study compares three deep learning models for Indian Sign Language (ISL) interpretation, with a focus on how well they learn spatial and temporal patterns using video-based gesture data. To capture detailed motion dynamics, MediaPipe’s Holistic model extracted 543 landmark sequences from video frames, including 21 right and left hands, 468 face, and 33 posture landmarks. The models tested include an Attention-Based BiDirectional Long Short-Term Memory (Bi-LSTM) Network, a Pose-Guided Graph Convolutional Network (PGCN), and a Spatiotemporal Transformer for Vector Sequences (STraVe). The findings show that the Attention-Based Bi-LSTM Network had the greatest classification accuracy of 95.87%, outperforming PGCN (94.70%) and STraVe (87.16%), capturing the complex spatiotemporal aspects of ISL.