A new method of data augmentation for machine learning using 3D model data is proposed. The method involves the use of STL data of an object to automatically generate a set of training data covering a continuous range of view angles and various backgrounds. It also involves the use of two CNN’s, one corresponding to the ‘object (parent class)’ and another to the ‘view angle (child class)’, that provides a two-stage classification to improve tolerance against over-classification. The performance of the method is demonstrated by comparing categorization results with conventional approach based on real-world photographs. The method shows satisfactory improvements over conventional method using photographed images.
Paper
Full text
Data Augmentation Based on 3D Model Data for Machine Learning
Semantic Scholar · Computer Science · 2019
Abstract
A new method of data augmentation for machine learning using 3D model data is proposed. The method involves the use of STL data of an object to automatically generate a set of training data covering a continuous range of view angles and various backgrounds. It also involves the use of two CNN’s, one corresponding to the ‘object (parent class)’ and another to the ‘view angle (child class)’, that provides a two-stage classification to improve tolerance against over-classification. The performance of the method is demonstrated by comparing categorization results with conventional approach based on real-world photographs. The method shows satisfactory improvements over conventional method using photographed images.