Scholarlink

<p>In recent years, there has been a lot of research in the fields of computer vision and multimedia to analyze human behavior and activities through images. A particular area of focus has been on estimating human pose, also known as skeleton estimation. Deep learning methods have been commonly used for this task, which primarily rely on the keypoint features of the human body. However, this approach can be limiting when there are occluded or incomplete poses, especially when multiple humans are present in a single frame. Other features like visibility conditions and body boundaries can also contribute to pose estimation in such cases. This paper outlines a method for multi-person pose estimation leveraging Sparse-RCNN with mixture models, aiming to reduce computational complexity while enhancing accuracy. The approach integrates innovative techniques of Sparse-RCNN such as learnable proposal boxes, dynamic heads, and an iteration structure to efficiently extract features with mixture models and human body masks. This approach has resulted in significant improvements, as indicated by an increase in average precision and faster processing compared to other state-of-the-art methods on the COCO and CIHP datasets.</p>

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

Novel Cauchy mixture modeling combined with the Sparse-RCNN architecture for enhanced multi-person pose estimation

  • 期刊论文

  • 发表日期：

    2025年6月27日

查看全文
  • Tahir Rizwan,
  • Yunze Cai,
  • Rhythm Vohra

摘要

In recent years, there has been a lot of research in the fields of computer vision and multimedia to analyze human behavior and activities through images. A particular area of focus has been on estimating human pose, also known as skeleton estimation. Deep learning methods have been commonly used for this task, which primarily rely on the keypoint features of the human body. However, this approach can be limiting when there are occluded or incomplete poses, especially when multiple humans are present in a single frame. Other features like visibility conditions and body boundaries can also contribute to pose estimation in such cases. This paper outlines a method for multi-person pose estimation leveraging Sparse-RCNN with mixture models, aiming to reduce computational complexity while enhancing accuracy. The approach integrates innovative techniques of Sparse-RCNN such as learnable proposal boxes, dynamic heads, and an iteration structure to efficiently extract features with mixture models and human body masks. This approach has resulted in significant improvements, as indicated by an increase in average precision and faster processing compared to other state-of-the-art methods on the COCO and CIHP datasets.