Sana Exercise Pose Landmark Dataset

Joint angle time series extracted from gym exercise demonstration videos using MediaPipe BlazePose (pose_landmarker_heavy). Released in support of the master's thesis "Real-Time Exercise Recognition and Repetition Counting System from Monocular Gym Video", Alma Mater Studiorum - Universita di Bologna, 2026. 330 CSV files, 117,658 rows, 16 exercise classes, ~60.5 MB. Format: CSV, 29 columns per row (frame index + 14 joint angles normalized to [0,1] + 14 visibility scores). Sampled at 10 FPS. No video, images, or biometric data. No personal identifiers. Single-subject tracking (num_poses=1): background individuals were never captured. Exercise classes (16): bicep_curl, crisscross, deadlift, donkey_kick, lateral_lunge, lateral_raise, leg_raise, lunge, other, plank, pushup, row, shoulder_press, situp, squat, step_up. Subjects: gym staff (personal trainers and employees) who demonstrated exercises voluntarily with full awareness of the recordings. Recorded at Sana HF gym, Bologna, Italy, Fall 2024 - Spring 2025. See README.md for full documentation, column descriptions, limitations, and citation instructions.