Shape-Centric Augmentation with Strong Pre- Training for Training from Rendered Images
Takeru Inoue, Masakazu Ohkoba, Kouji Gakuta, Etsuji Yamada, Aoi Kariya, Masakazu Kinosada, Yujiro Kitaide, Ryusuke Miyamoto · 2024
Acquiring purpose-specific data is crucial for applying deep learning to applications. However, creating such data can be labor-intensive, necessitating the development of datasets at minimal cost. Consequently, there has been a recent trend towards utilizing computer-generated imagery (CGI) to generate purpose-specific datasets. Nonetheless, models trained on CGI data often exhibit suboptimal performance on real images due to domain gap issues. Thus, bridging this gap between synthetic and real data domains is crucial. We are developing a real parts classification application using 3D CAD models as training data. This study discusses our strategies for addressing the domain gap challenge in this application. To mitigate this gap, we explore three primary approaches: selecting architectures with a strong shape bias, leveraging large-scale pre-training to enhance generalization performance, and enhancing the shape bias of classifiers by manipulating color characteristics in the training data. We present evidence demonstrating the effectiveness of these meth-ods in bridging the domain gap between CG I training data and real-world images in our part classification application: the top-l accuracy became 98.34 % at the best case.