Scholarlink

<p>Markerless motion capture has gained significant interest for its potential applications in sport andbiomechanics. The performing arts, and circus arts in particular, remains unexplored. This study evaluates the performance of three pose estimation models, MediaPipe, YOLOv8, and VitPose, on circus movements, and assesses their accuracy based on Mean Per Joint Position Error (MPJPE), Percentage of Correct Keypoints (PCK), and detection rate. Results indicate that VitPose significantly outperforms YOLOv8 and MediaPipe, with the lowest MPJPE (0.192&#xa0;m) and the highest PCK (68.1%). Fine-tuning VitPose with circus-specific training data further improved accuracy, with a MPJPE of 0.098&#xa0;m and a PCK of 78.7%. To better capture the unique challenge of circus arts, custom evaluation metrics that assess pose estimation performance in acrobatic movement, such as inverted position, leg opening, hip flexion and body velocity, were developed. All models showed substantial errors at inverted position and large leg angle. These findings highlight the limitations of current markerless pose estimation models for accurately tracking extreme circus movements. Future research should explore expanded data sets with less traditional movements or occlusion-aware deep learning techniques.</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

Evaluating MediaPipe, YOLOv8, and VitPose for dynamic circus motion analysis

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Marion Cossin,
  • Alexandre Laforest

摘要

Markerless motion capture has gained significant interest for its potential applications in sport andbiomechanics. The performing arts, and circus arts in particular, remains unexplored. This study evaluates the performance of three pose estimation models, MediaPipe, YOLOv8, and VitPose, on circus movements, and assesses their accuracy based on Mean Per Joint Position Error (MPJPE), Percentage of Correct Keypoints (PCK), and detection rate. Results indicate that VitPose significantly outperforms YOLOv8 and MediaPipe, with the lowest MPJPE (0.192 m) and the highest PCK (68.1%). Fine-tuning VitPose with circus-specific training data further improved accuracy, with a MPJPE of 0.098 m and a PCK of 78.7%. To better capture the unique challenge of circus arts, custom evaluation metrics that assess pose estimation performance in acrobatic movement, such as inverted position, leg opening, hip flexion and body velocity, were developed. All models showed substantial errors at inverted position and large leg angle. These findings highlight the limitations of current markerless pose estimation models for accurately tracking extreme circus movements. Future research should explore expanded data sets with less traditional movements or occlusion-aware deep learning techniques.