Scholarlink

Yoga is a discipline originating from ancient India, concentrating on overall well-being, incorporating physical, mental, and spiritual activities. When performing yoga postures without expert guidance, one must concentrate on form and technique to prevent injuries, which necessitates significant empirical study in this regard. The main objective of our study is to assess still image-based classification of yoga postures using five of the most popular deep learning models: Vision Transformer, DenseNet201, ResNet50, InceptionV3, and VGG19. These models are tested on two recently developed standard yoga pose classification datasets: one created by Niharika Pandit and the other by A. Mohan Kumar. The comparison results conclude that the InceptionV3 model showed the best classification accuracy of 98.29% on the Niharika dataset, while the DenseNet201 model showed the best accuracy of 87.54% on the other dataset. This in-depth examination highlights the relative strengths and weaknesses present in each model, helping to choose the best one for classifying yoga poses. This groundbreaking study validates the proposed method and marks a new era where high-quality yoga instructions are available to everyone, regardless of their location or financial status.

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

Performance Evaluation of Different Deep Learning Models for Automatic Yoga Pose Classification

  • 章节

  • 发表日期：

    2025年

查看全文
  • Agnish Paul,
  • Shuvam Ghosh,
  • Pawan Kumar Singh,
  • Jana Shafi,
  • Muhammad Fazal Ijaz

摘要

Yoga is a discipline originating from ancient India, concentrating on overall well-being, incorporating physical, mental, and spiritual activities. When performing yoga postures without expert guidance, one must concentrate on form and technique to prevent injuries, which necessitates significant empirical study in this regard. The main objective of our study is to assess still image-based classification of yoga postures using five of the most popular deep learning models: Vision Transformer, DenseNet201, ResNet50, InceptionV3, and VGG19. These models are tested on two recently developed standard yoga pose classification datasets: one created by Niharika Pandit and the other by A. Mohan Kumar. The comparison results conclude that the InceptionV3 model showed the best classification accuracy of 98.29% on the Niharika dataset, while the DenseNet201 model showed the best accuracy of 87.54% on the other dataset. This in-depth examination highlights the relative strengths and weaknesses present in each model, helping to choose the best one for classifying yoga poses. This groundbreaking study validates the proposed method and marks a new era where high-quality yoga instructions are available to everyone, regardless of their location or financial status.