3Dance is a tool designed to help users safely learn and practice dance by comparing their movements to instructor videos in real time. Using OpenPose pose estimation, it tracks body key points and provides accuracy feedback, achieving about 3% error compared to human instructors while operating at up to twice the original video speed. The system also delivers real-time feedback at 10–30 fps, making it practical for live practice. Developed as part of a 2025 academic project, it focuses on providing accessible, real-time dance training through motion analysis.
The program uses OpenPose with CNNs for real-time pose estimation and the Godot engine to display keypoint data in a lightweight, user-friendly interface. It calculates accuracy by comparing joint angles between a student and instructor using atan2, measuring percent differences, and averaging them across frames to produce a final grade. OpenPose 1.7.0 is chosen for its GPU optimization and efficient keypoint detection, while Godot 4.6 supports flexible UI design and 2D/3D integration with minimal overhead. The system also includes setup steps for processing keypoint and image data, handling invalid inputs, and enabling real-time performance through optimized configurations.
Testing shows the system can deliver real-time pose estimation feedback at about 10 fps with minimal latency (under 0.1 seconds), or up to 30 fps when image feedback is disabled. Preprocessed inputs and buffered image loading can further boost performance to roughly twice the original video speed, with potential for additional improvements using languages like C++ or Python. Evaluation with a dance professor found the grading system has an average error of about 3.03% and a 3.26% standard deviation compared to human assessment. Future work aims to reduce this error to around 2% by training OpenPose specifically on dance movements.