MACHINE PERCEPTION AND ROBOTICS GROUP

Dept. of Robotics Science and Technology,
Chubu University

Deep convolutional neural networks require a huge amount of data samples to train efficient networks. Al- though many benchmarks manage to create abundant samples to be used for training, they lack efficiency when trying to train convolutional neural networks up to their full potential. The data augmentation is one of the solutions to this problem, but it does not consider the quality of samples, i.e. whether the augmented samples are actually suitable for training or not. In this paper, we propose a method that will allow us to select effective samples from an augmented sample set. The achievements of our method were 1) to be able to generate a large amount of augmented samples from images with labeled data and multiple background images; and 2) to be able to select effective samples from the additionally augmented ones through iterations of parameter updating during the training process. We utilized exhaustless sample generation and effective sample selection in order to perform recognition and segmentation tasks. It obtained the best performance in both tasks when compared to other methods using, or not, sample generation and/or selection.