ALEX-GYM-1 : A Novel Dataset and Hybrid 3D Pose Vision Model for Automated Exercise Evaluation

Overview

ALEX-GYM-1 is a dataset and hybrid model designed for automated gym exercise evaluation. It combines synchronized multi-view videos (frontal and lateral) with detailed biomechanical annotations for exercises like Squats, Lunges, and Single-Leg Romanian Deadlifts. The dataset supports both pose-based and vision-based analysis, enabling accurate and efficient exercise quality assessment.

The hybrid model integrates:

• Vision-based features: Extracted using a 3D ResNet architecture.
• Pose-based features: Engineered from 3D skeletal keypoints.

This approach achieves high accuracy in classifying exercise performance and detecting errors.

Dataset Highlights

• Exercises: Squats (295 videos), Lunges (106 videos), Deadlifts (269 videos).
• Participants: 45 individuals (diverse age and gender groups).
• Annotations: Biomechanical criteria for each exercise.
• Data: Includes video frames, pose keypoints (JSON), and metadata (Excel).

Model Performance

The hybrid model outperforms single-modality approaches:

• Pose-based model: Focuses on joint relationships.
• Vision-based model: Captures spatio-temporal dynamics.
• Hybrid model: Combines both for superior accuracy.

Getting Started

1. Install Dependencies

pip install -r requirements.txt

2. Download Dataset

Download the dataset and organize it as follows:

ALEX-GYM-1/
├── Squat_frames/
├── Deadlift_frames/
├── Lunges_frames/
├── squat.xlsx
├── deadlift.xlsx
├── lunges.xlsx
├── front_pose_squat.json
├── front_pose_deadlift.json
├── front_pose_lunges.json
├── lat_pose_squat.json
├── lat_pose_deadlift.json
└──  lat_pose_lunges.json

Results

The hybrid model achieves:

• Squats: Hamming Loss 0.0259
• Deadlifts: Hamming Loss 0.0488
• Lunges: Hamming Loss 0.0756

Acknowledgment

This work is funded by the Science and Technology Development Fund (STDF), Egypt, under project ID 51399.

---

This work is accepted at the ICINCO 2025 conference. The authors are:

• Ahmed Hassan
• Abdelaziz Essam
• Ahmed Yasser
• Prof. Walid Gomaa