Optimizing Data Augmentation for Semantic Segmentation on Small-Scale Dataset

Abstract

Augment the training data has a significant effect for deep learning on small-scale dataset. For practical semantic segmentation applications, it is a hard work to collect and annotate enough training data for training the deep neural network. In this paper, we focus on which data augmentation (DA) method is better, and what combination of different DA methods can improve the network performance more. Our target application is highland AI-Ranch which is hard to collect many training data. We firstly collect and produce a small-scale open source of sheep segmentation dataset including hundreds images, referred to as SSG dataset. Seven frequently used data augmentation methods are evaluated, including global augmentation (augment for the whole image) such as flipping, and local augmentation (augment only for the region of interest) such as cropping, etc. Especially, a novel image compression global DA method is proposed which can achieve the best augmentation performance in global methods. Furthermore, we explore the performance of the cross-combination data augmentation technique when applying to a small-scale semantic segmentation dataset. As different DA method will cover the different sample distribution, more augmentation fed more good training data and meanwhile more bad training data to the network. Therefore, too much augmentation may pull down the performance sometimes. Experiment results show that the combination of compression, cropping and local shift can achieve the best augmentation performance for our AI-Ranch application, the average coverage mean-IoU improve from 73.3% to 91.3%, even better than the combination of whole augmentation methods.