Deep Optical Flow Supervised Learning With Prior Assumptions

Traditional methods for estimating optical flow use variational model that includes data term and smoothness term, which can build a constraint relationship between two adjacent images and optical flow. However, most of them are too slow to be used in real-time applications. Recently, convolutional neural networks have been used in optical flow area successfully. Many current learning methods use large data sets that contain ground truth for network training, which can make use of prior knowledge to estimate optical flow directly. However, these methods overemphasize the factor of deep learning and ignore advantages of many traditional assumptions used in variational framework for optical flow estimation. In this paper, inspired by classical energy-based optical flow methods, we propose a novel approach for dense motion estimation, which combines traditional prior assumptions with supervised learning network. During training, the variation in image brightness, gradient and spatial smoothness are embedded in network. Our method is tested on both synthetic and real scenes. The experimental results show that employing the prior assumptions during training can obtain more detailed and smoothed flow fields and can improve the accuracy of optical flow estimation.