Neural Network Human Motion Style Transfer via Object Attributes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current motion capture technologies require significant time, money, and manpower for recapturing or re-adjusting motion data when scenes or creative purposes change, and data-driven methods involve extensive manual preprocessing, leading to errors and inconsistent animation quality.
Innovation Solution
A generation method for a human body motion editing model using a neural network that takes initial motion sequences and target object attributes to generate new motion sequences, reducing manual intervention and improving data reuse by transferring motion styles based on object attributes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If motion capture technology is used to record real motion information, then smooth and natural character motion can be obtained, but when scene or creative purpose changes, recapturing or re-adjusting is required which wastes time and money
Solution Approach 1:
The patent changes the parameters of the motion data by using neural network models to transform initial motion sequences into target motion sequences based on object attributes. This allows the same captured motion to be adapted to different scenes and purposes by adjusting parameters like object type, interaction mode, and motion characteristics, eliminating the need for recapture.
Solution Approach 2:
The patent creates virtual copies of motion data through neural network generation. Instead of recapturing real motion, the system generates synthetic motion sequences that copy and adapt the characteristics of original captured motion to new scenarios, preserving the quality while avoiding repeated capture costs.
2Productivity
If data-driven motion synthesis is used, then motion can be generated without recapture, but extensive manual preprocessing including segmentation, alignment, and marking is required which introduces errors
Solution Approach 1:
The patent replaces the mechanical manual preprocessing system with an automated neural network system. The neural network automatically performs segmentation, alignment, and feature extraction without human intervention, eliminating manual errors while maintaining high productivity. The end-to-end learning framework substitutes manual operations with automated computational processes.
Solution Approach 2:
The neural network model performs self-service by automatically preprocessing the motion data without requiring manual segmentation, alignment, or marking. The system independently extracts features, aligns sequences, and generates target motion, ensuring consistency and quality while maintaining high efficiency.
3Ease of manufacture
If manual data preprocessing is performed for data-driven motion synthesis, then motion sequences can be processed, but errors at any stage lead to animation failure and smooth natural motion cannot be ensured
Solution Approach 1:
The patent substitutes manual preprocessing operations with automated neural network processing. The neural network maintains manufacturing precision by learning optimal preprocessing parameters from training data, ensuring accurate segmentation, alignment, and feature extraction without human error while preserving ease of manufacture through automation.
Solution Approach 2:
The patent implements feedback mechanisms where the neural network evaluates the quality of processed motion sequences and adjusts preprocessing parameters accordingly. This feedback loop ensures that segmentation, alignment, and feature extraction maintain high precision by continuously optimizing based on output quality metrics, preventing error propagation.
Data Source
AI summary
A generation method of a human body motion editing model, storage medium and electronic device. The method includes taking initial motion sequence, target object attributes and target motion sequence as training samples; inputting initial motion sequence and target object attribute into preset neural network model, obtaining generated motion sequence output by preset neural network model; training preset neural network model according to target motion sequence and generated motion sequence to obtain trained human body motion editing model. Through using intermotion motion sequence of person and object as training sample, human body motion style migration is realized only by utilizing object attributes, so human intervention amount in data preprocessing process is reduced, human-object intermotion movement matched with attributes can be correspondingly generated by applying different attributes of same object, and reuse value of motion capture data is improved.