3D Data Augmentation via Weighted Local Non-Rigid Transformation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

The challenge is to effectively augment 3D data for improving AI model performance with limited data, particularly addressing the difficulty in processing and securing diverse 3D data resources and preventing heterogeneity in new data through non-rigid transformations.

Innovation Solution

A method and apparatus that apply a non-rigid transformation based on kernel regression to local parts of 3D data, involving preprocessing, anchor point sampling, transformation parameter sampling, and calculating a non-rigid transformation matrix to generate diverse 3D data with minimal resources.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If non-rigid transformation is applied to local parts of 3D data, then data diversity is improved, but data heterogeneity increases

Engineering Contradiction:

Improvedata diversityVSAvoiddata heterogeneity

Core Design Contradiction:

Adaptability or versatilityVSStability of the object's composition

Solution Approach 1:

The patent applies different transformation parameters to different local regions of the 3D data by dividing the data into multiple regions and generating region-specific transformation matrices. This allows each local part to be transformed differently, improving data diversity while maintaining overall data consistency through the use of anchor points and weighted combinations.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent transforms 3D data by changing transformation parameters (rotation, scaling, translation) applied to different regions. By sampling multiple transformation parameters and combining them with weights based on anchor point distances, the system generates diverse transformed data while controlling heterogeneity through parameter interpolation.

Inventive Principle:

Principle #35Parameter changes

2Productivity

If 3D data is processed and transformed, then data augmentation is improved, but processing resources increase

Engineering Contradiction:

Improvedata augmentationVSAvoidprocessing resources

Core Design Contradiction:

ProductivityVSUse of energy by moving object

Solution Approach 1:

The patent segments the 3D data into multiple regions and processes each region independently with its own transformation parameters. This segmentation allows for efficient parallel processing and reduces the computational burden compared to transforming the entire dataset uniformly, thereby improving data augmentation efficiency while managing resource consumption.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent generates multiple copies of the original 3D data by applying different transformation parameters to create transformed versions. These copies are generated through mathematical transformations rather than physical duplication, which is computationally efficient and allows rapid data augmentation without proportionally increasing processing resources.

Inventive Principle:

Principle #26Copying

Data Source

PatentUS20240346790A13D data augmentation method using weighted local transformation and apparatus therefor

Publication Date: 2024.10.17 KOREA UNIV RES & BUSINESS FOUND

AI summary

The present disclosure relates to a three-dimensional (3D) data augmentation method using a weighted local transformation and an apparatus for the same. More specifically, the present disclosure relates to a method and apparatus of significantly augmenting 3D data required to improve the performance of an artificial intelligence model with only limited 3D data by applying a non-rigid transformation to local part(s) of the 3D data.