Transfer Learning and On-Fly Data Augmentation for Scene UnderstandingUsing InceptionResNet
Michael Daniel Nachipyangu, Jiangbin Zheng, Palme Mawagali · 2023
Deep learning models require large amounts of data to achieve good results. However, most datasets consist of images taken from similar angles, brightness levels, and orientations, which do not reflect the diverse reality of scenes. To address this issue, data augmentation techniques are employed to generate images that mimic actual scenarios, thereby increasing the training data for the model. In this paper, we propose an on-the-fly data augmentation approach that enhances the dataset while minimizing the need for additional storage by not saving augmented images to disk. We evaluate different pretrained and trained-from-scratch Convolutional Neural Network (CNN) models on benchmark scene datasets (Scene15 and MIT67), and our results demonstrate that fine-tuning the InceptionResNetV2 model achieves competitive performance compared to state-of-the-art methods on these datasets with accuracy of 95% and 86% respectively. This research contributes to creating more realistic scene representations through data augmentation while optimizing disk space usage. Furthermore, we highlight the effectiveness of data augmentation as a regularization technique by reducing loss. The findings presented in this paper provide valuable insights for scene understanding tasks and have implications for various applications such as education, healthcare systems, autonomous vehicles, and domestic robot navigation.