handkeypoint

Hand keypoint refers to a spatial coordinate that marks a specific anatomical landmark on the hand, such as the wrist or finger joints. In computer vision and human pose estimation, hand keypoints are used to describe hand pose and articulation from images or videos.

The most common convention uses 21 keypoints per hand, covering the wrist, the finger bases (metacarpophalangeal joints), the middle and end joints, and the fingertips. This standard is adopted by popular datasets and models, including MediaPipe Hands. Keypoints can be detected in 2D image coordinates or lifted to 3D using depth cues or multi-view geometry.

Methods for estimating hand keypoints include heatmap-based deep networks that predict per-keypoint likelihood maps, direct coordinate regression, and graph-based models that exploit the hand’s skeletal structure. For video, temporal models improve stability. 3D hand pose estimation often fuses depth data or learns from synthetic data; 2D methods are commonly extended with 3D lifting to obtain three-dimensional coordinates.

Applications of hand keypoint estimation include gesture recognition, sign language interpretation, hand tracking for augmented reality and virtual reality, human–computer interaction, and robotic manipulation. Challenges arise from occlusion between fingers, self-occlusion when hands interact, rapid articulation, variations in scale and lighting, and variability in annotation quality.

Public datasets provide hand keypoint annotations for benchmarking. Notable examples include FreiHAND, the Rendered Hand Pose Dataset (RHD), NYU Hand Pose, STB, and DexYCB. Evaluation metrics commonly used are PCK (percentage of correct keypoints), MPJPE (mean per-joint position error), and related area under the curve measures.