Summary
International Workshop on Smart Info-Media Systems in Asia
2023
Session Number:RS3
Session:
Number:RS3-2
Improving Classification Accuracy of Real Images by Style Transfer Utilized on Synthetic Training Data
Takeru Inoue, Youichi Tomita, Kouji Gakuta, Etsuji Yamada, Aoi Kariya, Masakazu Kinosada, Yujiro Kitaide, Ryusuke Miyamoto,
pp.71-76
Publication Date:2023/8/31
Online ISSN:2188-5079
PDF download
Summary:
In general, applying machine learning requires a significant effort to create a suitable training dataset that matches the intended purpose. To reduce such efforts, a framework is sometimes employed to automatically perform rendering of 3D models and assign class labels. However, as commonly known, classifiers trained directly on images rendered from 3D models tend to exhibit lower accuracy when applied to real-world images. In this study, we address this issue by proposing a method that emphasizes the recognition of object shapes during classifier training. By appropriately applying style transfer to the generated training dataset from 3D models, we were able to improve the classification accuracy on the real-world images dataset by up to 11%.