Hand Pose Estimation and Hand Tracking

Hand-pose estimation detects a hand and predicts locations of key points such as fingertips, joints, and wrist. MediaPipe Hand Landmarker describes a set of 21 landmarks for each detected hand and provides image and world-coordinate results. Its task can run on a still image, video, or live stream. In video or live-stream modes, the pipeline can use previous detections to localize hands in later frames, reducing repeated work when tracking continues smoothly. Landmarks are geometric estimates, not a full understanding of gesture meaning or intent. A pinched thumb and index finger could represent a control gesture in one app and an ordinary movement in another. Occlusion, motion blur, lighting, skin/background contrast, camera angle, and hands leaving the frame can reduce detection quality or cause identity swaps when two hands cross. Applications must define which landmark patterns mean an action and provide a way to pause or undo unintended commands. A practical test measures landmark error, detection misses, identity switches, latency, and user comfort across representative people and environments. If used for sign language or clinical measurement, a small set of hand points alone is not a translation or diagnosis. Build task-specific labels and evaluate with the people and conditions expected in deployment. MediaPipe is one toolkit example; model size, coordinates, runtime, and behavior can differ across versions and other hand-tracking platforms.