A Multi-View Raw Video Dataset of Seven Fitness Exercises with Good/Bad Form Labels

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.