Point Cloud Deformation Composition for GAN Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for generating training data using generative adversarial networks (GAN) are limited to images and struggle to apply to point cloud data, and methods for pasting extracted objects onto images do not effectively generate deformations, making it difficult to extend machine learning or deep learning techniques for point cloud data generation.
Innovation Solution
A deformation composition data generation apparatus and method that acquires and composes the distribution of displacement for points in point cloud data, allowing for the synthesis of deformations on planar and geometric structures using a deformation feature database and composition units.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If generative adversarial networks (GAN) are used to generate training data, then training data diversity is improved, but the technique is limited to images and cannot be applied to point cloud data
Solution Approach 1:
The patent copies the successful approach of GAN-based data augmentation from image processing and adapts it to point cloud data by using a similar generative framework (PointFlow) that learns the distribution of point cloud data and generates synthetic samples, thereby extending the applicability of generative models to 3D point cloud domains
Solution Approach 2:
The patent changes the parameter space from 2D image coordinates to 3D point cloud coordinates, adapting the generative model to handle the different data structure and dimensionality of point cloud data while maintaining the core generative adversarial framework
2Ease of manufacture
If extracted objects are pasted onto images to generate training data, then data generation is simplified, but deformations cannot be effectively generated
Solution Approach 1:
The patent replaces the mechanical object-pasting approach with a learned generative model (PointFlow) that uses neural networks to synthesize deformations, substituting manual geometric manipulation with data-driven probabilistic modeling to achieve both ease of generation and high precision
Solution Approach 2:
The generative model learns to generate deformations autonomously by training on point cloud data with various deformations, enabling the system to self-generate diverse deformation patterns without manual intervention or pre-defined deformation templates
Data Source
AI summary
In the deformation composition data generation apparatus which is implemented by one or more processors, an acquisition unit acquires a distribution of amount of displacement for points in point cloud data, and a composition unit composes the amount of displacement to the points in the point cloud data according to the distribution.