A Multi-View Raw Video Dataset of Seven Fitness Exercises with Good/Bad Form Labels
Description
This is a raw video dataset for vision-based exercise analysis and posture quality analysis. The dataset consists of recordings of seven exercises: standing dumbbell side bend (abs), bent-over dumbbell row (back), alternating dumbbell bicep curl, standard push-up, standing dumbbell shoulder press, bodyweight squat, and overhead dumbbell triceps extension. The videos were recorded by going to the gym. Inside the gym the videos were recorded using smartphones. These videos were captured from three different views i.e. front, side, diagonal and was labeled by subject identity, exercise class, posture quality (good/bad). This data was collected from 26 subjects under real indoor conditions with natural variation in lighting and background clutter. The data was organized using a fixed naming convention encoding subject_id, exercise, form_quality, and view. This repository provides raw MP4 videos and the dataset can be used for extracting various data points in a csv format for pre-processing and also model training purposes. This resource is intended for people who are interested in computer vision, human movement analysis, digital health, and AI-assisted fitness systems.
Files
Steps to reproduce
This dataset was collected by recording seven fitness exercises from multiple participants using multiple camera viewpoints. Each video was manually reviewed and categorized according to the exercise performed, execution quality (good or bad), viewing angle (front, side, diagonal), and participant ID. The original videos were organized and verified to ensure consistent file naming. Three alternative folder organizations are provided to support different research workflows: 1. Dataset Subject-wise – Videos grouped by participant, then exercise, quality, and camera angle. 2. Dataset Exercise-wise – Videos grouped by exercise, then quality, camera angle, and participant. 3. Dataset Exercise-Quality – Videos grouped by exercise, quality, and camera angle for quick access during model training. To reproduce the dataset structure: 1. Download and extract the dataset archive. 2. Select the folder organization that best fits your research objective. 3. The video files are identical across all three organizations; only the folder hierarchy differs. 4. Use the videos directly for computer vision, machine learning, human pose estimation, exercise recognition, or posture/form classification tasks. 5. File names uniquely identify the participant, exercise, execution quality, and camera viewpoint, enabling straightforward parsing and automated processing.