Scholarlink

To build systems for device control, entertainment, and human-machine interaction using hand gestures, the hand data area needs to be detected, hand posture needs to be estimated, and hand gestures need to be recognized. Hands need to be done quickly and accurately. Therefore, each step in the system construction model needs to be performed quickly and accurately. The advent of deep learning (DL) has brought very impressive results in solving computer vision problems, especially CNN for object detection (OD) and recognition problems. From there, DL is a good approach to quickly and accurately solve the hand detection (HDe) problem. In this study, we fine-tuned the right-HDe model using the YOLOv8 (YOLOv8n, YOLOv8m, YOLOv8l, and YOLOv8x) on the HOI4D dataset, which is the latest DL network of the YOLO family for OD. Right-HDe results on YOLOv8x are the best (P = 99.2%, R = 98.81%, mAP@50 = 99.4%). The HDe results must be carried out for each hand action of the HOI4D dataset. At the same time, the computation time on the GPU and CPU of YOLOv8 is (YOLOv8n = 197 fps, YOLOv8m = 194 fps, YOLOv8l = 192 fps, YOLOv8x = 188 fps) and (YOLOv8n = 1.73 fps, YOLOv8m = 1.64 fps, YOLOv8l = 1.62 fps, YOLOv8x = 1.55 fps), respectively.

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

Fast and Accurate Right-Hand Detection Based on YOLOv8 from the Egocentric Vision Dataset

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Van-Dinh Do,
  • Trung-Minh Bui,
  • Van-Hung Le

摘要

To build systems for device control, entertainment, and human-machine interaction using hand gestures, the hand data area needs to be detected, hand posture needs to be estimated, and hand gestures need to be recognized. Hands need to be done quickly and accurately. Therefore, each step in the system construction model needs to be performed quickly and accurately. The advent of deep learning (DL) has brought very impressive results in solving computer vision problems, especially CNN for object detection (OD) and recognition problems. From there, DL is a good approach to quickly and accurately solve the hand detection (HDe) problem. In this study, we fine-tuned the right-HDe model using the YOLOv8 (YOLOv8n, YOLOv8m, YOLOv8l, and YOLOv8x) on the HOI4D dataset, which is the latest DL network of the YOLO family for OD. Right-HDe results on YOLOv8x are the best (P = 99.2%, R = 98.81%, mAP@50 = 99.4%). The HDe results must be carried out for each hand action of the HOI4D dataset. At the same time, the computation time on the GPU and CPU of YOLOv8 is (YOLOv8n = 197 fps, YOLOv8m = 194 fps, YOLOv8l = 192 fps, YOLOv8x = 188 fps) and (YOLOv8n = 1.73 fps, YOLOv8m = 1.64 fps, YOLOv8l = 1.62 fps, YOLOv8x = 1.55 fps), respectively.