Dance Pose Estimation via Deep Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current human pose estimation models are not designed for professional dance evaluation and struggle with recognizing exaggerated body movements and overlapping actions, leading to inaccurate pose recognition.
Innovation Solution
A professional dance evaluation method using Deep Transfer Learning that combines Transfer Learning principles with pose feature training to build a Human Pose Estimation model, utilizing preprocessed videos to obtain and compare keypoint data for evaluation, optimizing recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If general human pose estimation models (openpose, alphapose, deeppose) are used for dance evaluation, then the models can process normal scenarios with good recognition effects, but they cannot accurately recognize professional dance poses with exaggerated movements and body overlap
Solution Approach 1:
The patent applies preliminary action by pre-processing dance video data to extract skeleton information and generate training datasets specifically tailored for professional dance poses. This preparatory work enables the model to better handle exaggerated movements and body overlaps before actual evaluation occurs.
Solution Approach 2:
The patent changes key parameters including loss function weights (emphasizing keypoint detection accuracy), model architecture (adding specific layers for dance pose features), and training data composition (using professionally annotated dance videos). These parameter adjustments transform a general pose estimation model into one specialized for professional dance evaluation.
2Productivity
If multiple students perform consecutively in examinations, then comprehensive evaluation can be conducted, but it becomes difficult and tiresome for teachers to visually observe all dancing action details
Solution Approach 1:
The patent replaces the mechanical system of human visual observation with an automated computer vision system. The deep learning model processes video inputs to extract pose information, substituting the teacher's visual observation capability with algorithmic analysis that does not fatigue and can process multiple students continuously.
Solution Approach 2:
The system enables self-service evaluation where the examination system automatically processes student performances without requiring continuous teacher intervention. The model independently analyzes each student's pose data and generates evaluation results, freeing teachers from the tedious task of observing every detail of each performance.
3Measurement precision
If deep transfer learning is used to build a specialized Human Pose Estimation model for dance, then recognition accuracy for professional dance poses is improved, but model complexity and training requirements increase
Solution Approach 1:
The patent achieves universality by building upon pre-trained deep learning models that have learned general human pose estimation capabilities. This foundation allows the specialized dance model to leverage existing knowledge while adding dance-specific features, reducing the need to train from scratch and managing complexity through transfer learning.
Solution Approach 2:
The patent applies preliminary action through pre-training on general pose estimation data before fine-tuning on dance-specific data. This staged approach allows the model to first learn fundamental human body structures and movements, then specialize for dance poses, reducing overall training complexity and computational requirements.
Data Source
AI summary
The present invention provides a professional dance evaluation method for implementing Human Pose Estimation based on Deep Transfer Learning. First of all, the Transfer Learning principle of deep learning is combined with the pose features of professional dance training to build a Human Pose Estimation model. Afterwards, the video of demonstration dancing actions is collected and imported into the Human Pose Estimation model to obtain the time-dependent body keypoint data as the reference standard for evaluation. Finally, the video of the examinee's dancing actions is collected and imported into the Human Pose Estimation model to obtain the body keypoint data of the examinee's dancing actions, the similarity between it and the reference standard for evaluation is used for evaluating the standard level of dancing pose.