About Fitness-AQA

Fitness-AQA is an AI-based system for assessing workout form and detecting posture errors during exercise, introduced at ECCV 2022. Designed for real-world gym scenarios, it analyzes video to evaluate exercise quality even under challenging conditions such as difficult camera angles, varied clothing, occlusion from gym equipment, and other factors that typically cause standard pose estimators to fail. The system currently targets three exercises: BackSquat, OverheadPress, and BarbellRow. It employs multiple self-supervised representation learning methods, including a pose contrastive learning framework, a motion disentangling approach, and a pose and appearance disentangling technique, as well as a quasi-synchronization method for handling in-the-wild video footage. The project also introduces the Fitness-AQA dataset, the largest fine-grained exercise action quality assessment dataset, available for non-commercial research upon request. Use cases include AI fitness coaching, injury prevention through form cor

p

Published by

paritoshparmar

Visit View Profile

README.md

View on GitHub

📢 Announcement: Workshop on Action Quality Assessment at CVPR 2026! 📢

Jan 2026 📢 📢 📢 We will be organizing Workshop on Skilled Activity Understanding, Assessment and Feedback Generation (SAUAFG) at CVPR 2026! More info on SAUAFG Website. Consider submitting your papers! See you in Denver, Colorado!

Fitness-AQA (Fitness Action Quality Assessment) [ECCV'22]

Full paper: Domain Knowledge-Informed Self-Supervised Representations for Workout Form Assessment

Introduction

Analyzing a person's posture during exercising is necessary to prevent injuries and maximizing muscle mass gains. In this work, we present an AI-based approach to detect errors in workout form. Our approach is particularly applicable/useful in real world gym scenarios, where off-the-shelf pose estimators fail to effectively capture person's pose due to challenging factors like camera recording angles, clothing styles, occlusions from gym equipment, etc. We applied our system to detect posture errors in three exercises: 1) BackSquat; 2) OverheadPress; and 3) BarbellRow. For this we collected the largest fine-grained exercise action quality assessment dataset, Fitness-AQA. Details on our self-supervised representation learning approaches and dataset are as follows:

Our Self-Supervised Pose Contrastive Learning Approach for Fine-grained Action Assessment

cvcspc

Our Self-Supervised Motion Disentangling Approach for Fine-grained Action Assessment

motion_disentangling

Our Self-Supervised Pose and Appearance Disentangling Approach

pose_appearance_disentangling

Our Method for Synchronizing In-the-Wild Videos

video_quasi_syncing_technique

Fitness-AQA Dataset for Fine-grained Exercise Action Quality Assessment

fitness-aqa_dataset

Dataset available from: https://forms.gle/PbPTX1eVxGpa3QG88. Please send us the request to access the dataset using this form. By requesting the dataset, you agree to the terms and conditions of usage. This dataset shall only be used for non-commercial purposes. Please check your spam folder in case you seem to not have received the access after requesting it; or please contact me if you are still have not received the access. Thank you!

Please feel free to reach out to me if you have any questions or face any problems.

If you find our work useful, please consider citing:

@inproceedings{parmar2022domain,
  title={Domain Knowledge-Informed Self-supervised Representations for Workout Form Assessment},
  author={Parmar, Paritosh and Gharat, Amol and Rhodin, Helge},
  booktitle={Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XXXVIII},
  pages={105--123},
  year={2022},
  organization={Springer}
}

🚀 Also Check Out Our New Approach! 🚀

Oct 2024: We have developed a new approach, NeuroSymbolic AQA, that builds upon this approach, but also analyses and scores using Professional Rules-based programs. It is Comprehensive and Explainable AQA which can generate Full Performance Reports for Actionable Insights!!! We encourage you to checkout [Code, Rules-based Programs, Dataset] [Demo] [Full Paper]

Fitness-AQA