A Fast And Accurate Scene Flow Estimation Strategy Based On Motion Decomposition

Scene flow is one of the most fundamental properties of scene,which gains increasing attention in computer vision.Large computational cost has limited scene flow from practical usage through the years.Meanwhile,the error propagation during multi-task estimation limits the accuracy of scene flow estimation.This paper briefly introduces the background of scene flow estimation,and discusses the relationship and the difference between the scene flow and other motion fields.On the basis of the sensor input,modeling concepts and solver choices,this paper analyzes the changes of scene flow estimation methods over the years,and categorizes the main strategies for scene flow estimation.The comparison between each kind is presented,while the pros and cons for each kind of method are illustrated.On the basis of the thorough categorization of most scene flow estimation methods and its developing progress,this paper introduces a novel framework for scene flow estimation,which decomposes the scene into multiple moving object regions and a non-moving background region.In the non-moving region,this paper leverage the camera pose estimation methods for structure-from-motion(SFM)problem and the simultaneous localization and mapping(SLAM)problem.The scene flow estimation problem in this region is turned into an over-determined least-square problem,which significantly reduces the computational cost.RANSAC is leveraged for outlier extraction,while sampling strategy is performed to lower the computational cost.When it comes to moving object regions,a sparse-to-dense interpolation strategy is performed.This paper introduces a novel three-dimensional geodesic distance,which better illustrates the intimacy between pixels.To verify the robustness of this method,this paper performs a basic test to verify the performance of the interpolation method and the camera pose estimation method.Firstly,this paper discusses the limitation of current manners to evaluate the performance of each scene flow estimation method.Detailed evaluation protocols,error metrics,and publicly available datasets are presented.Afterwards,the paper performs parameters adjustment and module decision on the basis of the average EPE on each training set.A thorough quantitative evaluation on each kind of publicly-available dataset is performed with the parameters and modules decided,along with the qualitative evaluation for visualization.The evaluationresults demonstrate the robustness,effectiveness,and accuracy of the proposed method,which achieves top-tier performance concerning accuracy and efficiency.