Data augmentation methods for semantic segmentation-based mobile robot perception system

Abstract

Data augmentation has become a standard technique for increasing deep learning models? accuracy and robustness. Different pixel intensity modifications, image transformations, and noise additions represent the most utilized data augmentation methods. In this paper, a comprehensive evaluation of data augmentation techniques for mobile robot perception system is performed. The perception system based on a deep learning model for semantic segmentation is augmented by 17 techniques to obtain better generalization characteristics during the training process. The deep learning model is trained and tested on a custom dataset and utilized in real-time scenarios. The experimental results show the increment of 6.2 in mIoU (mean Intersection over Union) for the best combination of data augmentation strategies.