Research On Pixel Flow Estimation Deep Models Based On Flow Boundaries Detection

In recent years,with the rapid development of computer vision,some emerging fields,such as autonomous cars,augmented reality and virtual reality,have achieved a great deal.However,as the industry advances,these fields make further requirements in accuracy as well as efficiency.Two important subtasks,i.e.optical flow estimation and binocular stereo matching,are still a focus in present research.Convolution neural network has achieved a huge breakthrough in some computer vision tasks,including image classification,object detection,semantic segmentation and so on.Following these works,deep learning is introduced to optical flow estimation and binocular stereo matching.Nevertheless,the related work is still at the beginning stage.The key of these two tasks lies in the construction of the correspondences between the dense pixels of image pairs,and the motion vectors of each pixel in time domain and space domain are calculated respectively(called flow in this thesis).I'm the first to define the two tasks uniformly as pixel flow estimation.The optical flow estimation focuses on the horizontal and vertical deviation displacement between two sequential pictures in time domain,while stereo matching focuses on the vertical deviation displacement of left and right image.I found that estimation errors usually occur at flow boundaries position.Yet in most cases,existing methods are based on object boundaries and are not effective for the flow boundaries.Thus,it is particularly important to obtain accurate flow boundaries and introduce it into deep model to solve the boundaries errors of flow estimation.The main work is as follows:1)I firstly propose a deep model called FBDNet for flow boundaries detection.My model proves deep model can learn the motion boundaries or depth boundaries from image pairs.FBDNet is obviously superior to traditional methods in terms of accuracy and efficiency.2)I propose a novel multi-task deep learning framework called MBANet for the boundaries detection of movement and optical flow estimation.My method leverages a two-stream decoder network for two tasks simultaneously.And I design an alternate and iterative algorithm.Experimental results show that different tasks provide complementary information and promote each other.My model can learn optical flow with more legible boundaries.It is worth mentioning that MBANet is a real-time method and dramatically boosts performance of existing real-time methods.3)I propose a multi-stage deep learning framework called DBANet for depth bound-aries detection and stereo matching.In the first stage,I conduct depth boundaries detection.In the second stage,I conduct disparity estimation with the detected depth boundaries and the original image pair as the input.The model shares the priori information of depth boundaries to the disparity estimation,which plays a guiding role.DBANet obtains more precise results and beats the famous DispNet in some datasets.