Hamish Pratt, Bernard Evans, Thomas Rowntree, Ian Reid, Steven Wiederman
Convolutional neural networks (CNNs) have become the-state-of-the-art for image classification and object detection tasks, as they have the ability to combine appearance features in a scene. CNNs used for detection and classification tasks primarily process single static images to combine the features. In a manner similar to biological brains, some neural networks also utilise motion as complementary information to aid object detection tasks. However, unlike the brain, these networks rarely classify ‘moving objects’ in a scene. Our research analyses a neural network’s ability to detect unique motion cues in scenes without any appearance, to understand the limits for neural networks to process motion information. We generated variant CNN models to understand different architectures that can process motion information and built a recurrent CNN with information skip layers for our experiments. By comparing our network’s detection rates against psychophysical stimuli used in human experiments, we found the neural network and humans both struggled to correctly detect unique motion in similar conditions. When trained for detecting higher orders of motion, stimuli observable by even small insects, the network responded strongly to the order of motion for which it was trained against, and was, for the majority, unresponsive to the other motion orders. To further test the ability of motion detection in neural networks, we trained a neural network against detecting repeating spatio-temporal signals inside a scene of random noise. The results from our experiments show that alongside convolutional neural networks' success in detecting appearance features for object classification, they are able to detect motion without appearance. With the understanding of similarities to biological brains and limitations in which these neural networks perform fundamental vision tasks like motion detection, we will have a better understanding of a network’s suitability for real-world applications. Presentation Link: meet.google.com/mob-nahz-rpa
