MyDeadlift Dataset: Labeled Deadlift Videos and Biomechanical Angle Data

Description

The MyDeadlift dataset contains 191 labeled videos of deadlift repetitions performed by seven male participants with different anthropometric characteristics and resistance-training experience. The recordings comprise three movement classes: Correct Movement (Gerakan Benar; GB; 83 videos), Rounded Back (Punggung Bungkuk; PB; 54 videos), and Knees Beyond Toes (Lutut Lebih Jari Kaki; LLK; 54 videos). Each video represents one complete repetition and was recorded from a fixed side-view position using the rear camera of an Apple iPhone 13 at 1920 × 1080-pixel resolution and 30 frames per second. The dataset is divided using a subject-wise splitting protocol. The training subset contains 131 videos from five participants, while the testing subset contains 60 videos from two participants, with 20 testing videos for each movement class. The repository also provides a processed table containing 21,813 frame-level records. Each record includes the data split, movement label, source video, frame number, knee angle, hip angle, and back-to-vertical angle. These biomechanical angles were calculated from body keypoints extracted using MediaPipe Pose. The dataset can be reused for human pose estimation, video-based exercise assessment, biomechanical movement analysis, human activity recognition, and the development or benchmarking of automated deadlift-technique classification methods.

Steps to reproduce

1. Recruit participants with resistance-training experience and record anonymized metadata: age, sex, height, body weight, training experience, and relevant fitness certification. Explain the research use of the recordings and obtain consent before recording. 2. Ask each participant to complete a 15–20-minute warm-up consisting of light cardiovascular activity, dynamic mobility exercises, bodyweight squats, and submaximal deadlift sets. 3. Record the repetitions using the rear camera of an Apple iPhone 13 at 1920 × 1080 pixels and 30 frames per second. Position the camera at the participant’s side to obtain a sagittal-plane view, at an approximate distance of 2.5 m and a height of 122.5–129.5 cm. 4. Record three classes: Correct Movement, Rounded Back, and Knees Beyond Toes. Maintain the same plate load across classes for each participant, adapted to individual capacity. 5. Segment the recordings so that each MP4 file contains one complete repetition, including the starting position, lifting phase, lockout, and lowering phase. Remove inactive segments before and after the repetition. 6. Inspect each video for body visibility, camera stability, repetition completeness, and consistency with its assigned class. Retain videos that adequately display the shoulder, hip, knee, and ankle regions and represent the intended movement. 7. Apply video-level labeling and name each file using the class–subject–repetition convention. For example, GB_S01_R01.mp4 represents Correct Movement, Subject 01, Repetition 01. The codes PB and LLK represent Rounded Back and Knees Beyond Toes. 8. Apply subject-wise splitting. Assign Subjects 01, 02, 03, 05, and 06 to the training subset and Subjects 04 and 07 exclusively to the testing subset. 9. Read each video sequentially and index every extracted frame, beginning with Frame 1. Preserve the temporal order, source video, data split, and movement label. 10. Process each frame using MediaPipe Pose to obtain two-dimensional coordinates for the shoulder, hip, knee, and ankle keypoints. 11. Calculate the knee angle from the hip–knee–ankle keypoints and the hip angle from the shoulder–hipknee keypoints. Calculate the angle formed by points u, v, and p, with v as the vertex, using θ = (180/π)[atan2(Yp−Yv, Xp−Xv) − atan2(Yu−Yv, Xu−Xv)]. Convert reflex angles greater than 180° to interior angles using 360°−θ. Calculate the back-to-vertical angle from the orientation of the hipshoulder vector relative to the vertical axis. 12. Store the results in a table containing Split, Label, Video, Frame, Sudut_Lutut, Sudut_Pinggul, and Sudut_Punggung_Vertikal. 13. Verify selected angle outputs using manual trigonometric calculations. Check the final table for missing values, duplicate rows, and duplicate video–frame combinations. Organize the videos, extracted-angle table, metadata, and documentation according to the repository structure.

Institutions

• Duta Wacana Christian University

  Yogyakarta, Yogyakarta

Categories

Licence