Benchmarking Human Mesh Estimation Methods for Physical Exercise Correction: A Quantitative and Qualitative Evaluation

^{Some examples from AVAFit dataset. It is a synthetic dataset focused on physical exercise, comprising SMPL-X meshes synchronized with RGB information from four viewpoints.}

AVAFIT is a fitness-oriented benchmark dataset for evaluating 3D human body reconstruction and pose estimation methods in exercise scenarios.

Dataset Overview

AVAFIT includes:

• 7200 exercise videos from 4 viewpoints.
• 60 exercises grouped into 10 categories.
• 3 executions of each exercise.
• Ground truth for each frame regarding SMPL-X meshes, depth, and binary masks.

The dataset is organized around:

• Participants
• Exercises
• Camera viewpoints
• Repetitions

Sample Videos

Below are example samples from the AVA-FIT dataset.

Action: A006 - Push-ups Subject: 003_lironghui Repetition: 3 Viewpoint: View 1 Zone: Aerobic Zone	Action: A057 - Jumping Jacks Subject: 001_herong Repetition: 2 Viewpoint: View 0 Zone: Machines Zone	Action: A047 - Side-lying Right Leg Backward Kick Subject: 004_aoyang Repetition: 3 Viewpoint: View 1 Zone: Pool Zone
Action: A042 - Sit-ups Subject: 006_liuyong Repetition: 3 Viewpoint: View 2 Zone: Press Zone	Action: A051 - Squat Jump Subject: 007_shenxiangwei Repetition: 2 Viewpoint: View 3 Zone: Press Zone	Action: A031 - Half Roll Back Subject: 008_tangjingyi Repetition: 1 Viewpoint: View 2 Zone: Pool Zone

Evaluation Code

The official evaluation code will be released soon.

Repository Status

This repository is under active development. Dataset access links, sample videos, documentation, and the official evaluation pipeline will be added progressively.

Citation

Citation information will be provided upon release.

License

License information will be provided upon release. License information will be provided upon release.