Distortion-Based Transparency Detection Using Deep Learning on a Novel Synthetic Image Dataset

Transparency detection is a hard problem, as suggested by animals and humans flying or running into glass. However, humans seem to be able to learn and improve on the task with experience, begging the question, whether computers are able to do so too. Making a computer learn and understand transparency would be beneficial for moving agents, such as robots or autonomous vehicles. Our contributions are threefold: First, we conducted a perception study to obtain insights about human transparency detection methods, when borders of transparent objects are not visible. Second, based on our study insights we created a novel synthetic dataset called DISTOPIA, which focuses on the warping properties of transparent objects, placed in a variety of natural scenes and contains over 140 000 high resolution images. Third, we modified and trained a deep neural network classification model with an attention module to detect transparency through warping. Our results show that a neural network trained on synthetic data depicting only distortion effects can solve the transparency detection problem and surpasses human performance.