Scholarlink

To build a support system and guide blind people to grasp objects intuitively, the hand pose, hand opening size, and grasping direction need to be determined accurately. To do this, the 3D hand pose needs to be estimated accurately and quickly to suit the requirements of blind people in reality, especially when the camera is mounted on the person’s chest/EVC and data is obscured by other objects and gaze directions. In the study, we fine-tune some 3D right-hand skeleton estimation (3D-R-HSE) models based on high-performance CNNs such as P2PR PointNet (PPN), Hand PointNet (HPN), V2V-PoseNet (VPN), and HandFoldingNet (HFN). The model fine-tuning is performed on the HOI4D dataset. CNNs to regress 3D hand pose/3D hand keypoints/3D hand skeleton from hand region data obtained from EVC. HOI4D is a benchmark dataset published at CVPR2022. 3D-HSE estimation results of PPN, HPN, VPN, and HFN with Erra are 32.71 mm, 35.12 mm, 26.32 mm, and 20.49 mm, respectively. HandFoldingNet has the best results and can be applied to further studies such as determining hand shape/type and size for object grasping or hand activity recognition. 3D HSE best result when using HFN on cam4th data, distance error is 12.41 mm. This helps us choose a good method for 3D-HSE. The time to estimate 3D HS is 5.4 fps.

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

Fine-Tuning of 3D Hand Pose Estimation on HOI4D Dataset by Convolutional Neural Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Van-Dinh Do,
  • Van-Hung Le

摘要

To build a support system and guide blind people to grasp objects intuitively, the hand pose, hand opening size, and grasping direction need to be determined accurately. To do this, the 3D hand pose needs to be estimated accurately and quickly to suit the requirements of blind people in reality, especially when the camera is mounted on the person’s chest/EVC and data is obscured by other objects and gaze directions. In the study, we fine-tune some 3D right-hand skeleton estimation (3D-R-HSE) models based on high-performance CNNs such as P2PR PointNet (PPN), Hand PointNet (HPN), V2V-PoseNet (VPN), and HandFoldingNet (HFN). The model fine-tuning is performed on the HOI4D dataset. CNNs to regress 3D hand pose/3D hand keypoints/3D hand skeleton from hand region data obtained from EVC. HOI4D is a benchmark dataset published at CVPR2022. 3D-HSE estimation results of PPN, HPN, VPN, and HFN with Erra are 32.71 mm, 35.12 mm, 26.32 mm, and 20.49 mm, respectively. HandFoldingNet has the best results and can be applied to further studies such as determining hand shape/type and size for object grasping or hand activity recognition. 3D HSE best result when using HFN on cam4th data, distance error is 12.41 mm. This helps us choose a good method for 3D-HSE. The time to estimate 3D HS is 5.4 fps.