Motion Generation Training With Relevance-Guided Regularization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in training a motion generation model without causing over-training, especially when only a small amount of motion data is available.
Innovation Solution
A training device that includes a motion data acquisition unit, a first generation unit, a determination unit, a relevance calculation unit, a regularization loss calculation unit, and an adversarial training processing unit, which utilize a first generation model, a determination model, and a regularization loss to train the motion generation model without over-training, even with a small amount of data, by combining source motions and calculating relevance and regularization losses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If cross-domain training is used to train a motion generation model with a small amount of motion data, then over-training is prevented and training efficiency is improved, but the method cannot be applied to Conditional-GAN architectures
Solution Approach 1:
The patent modifies the loss function parameters by adding a domain discrimination loss term to the standard Conditional-GAN loss function. This parameter change enables the loss function to simultaneously optimize for both motion generation quality and domain distinction, making cross-domain training applicable to Conditional-GAN architectures while preventing over-training
2Measurement precision
If a large amount of motion data is collected to train a recognition model with high accuracy, then recognition accuracy is improved, but the time and effort required for data collection increases significantly
Solution Approach 1:
The patent uses a motion generation model to synthesize pseudo motion data that copies the statistical characteristics and diversity of real motion data. This synthetic data copying approach enables training with large volumes of data without the time-consuming collection process, while maintaining the diversity needed for high recognition accuracy
3Quantity of substance
If pseudo motion data is generated to extend motion data for training, then data quantity is increased, but over-training occurs when using a small amount of actual motion data
Solution Approach 1:
The patent introduces a domain discriminator as an intermediary component that distinguishes between real and generated motion data. This intermediary enables the system to maintain training reliability by providing a mechanism to detect and balance the use of pseudo data, preventing over-training while still benefiting from data extension
Data Source
AI summary
A training device including a motion data acquisition unit that acquires first motion data related to a target motion, a first generation unit that generates pseudo first motion data by using a first generation model, a determination unit that calculates a determination loss indicating a degree of deviation between the first motion data and the pseudo first motion data using a determination model, a relevance calculation unit that reconfigures the target motion by a combination of basis motions and calculate a degree of relevance between the target motion and the basis motions, a regularization loss calculation unit that calculates a regularization loss indicating a degree of deviation between motion data related to the basis motions and the pseudo first motion data, and an adversarial training processing unit that adversarially trains the first generation model and the determination model using the determination loss and the regularization loss.