Effectiveness of Data Augmentation for CNN-Based Pupil Center Point Detection
Naoyuki Kan, Nagisa Kondo, Warapon Chinsatit, Takeshi Saitoh · 2018
Convolutional neural network (CNN) is one of the appearance-based methods, and this method has a potential for pupil center point detection problem. However, CNN requires a very large amount of training data. To overcome this problem, data augmentation (DA) is useful solution. In this paper, we investigate the effects of three DA approaches; affine transformation-based approach, synthetic image-based approach, and generative adversarial network (GAN)-based approach, and evaluate the effectiveness of DA for pupil center point detection task using an off-the-shelf CNN model. We set eleven conditions of the training data, and evaluated the detection accuracy. The result showed that higher detection accuracy was obtained for eye image by applying affine transformation. In our pupil center point detection task, it is found that affine transformation-based approach of DA is effective.