Simple method for generating evaluation data for scene flow algorithms. Issue 1 (1st January 2019)

• Record Type: Journal Article
• Authors: Jang, I.Y.; Lim, H.S.; An, S.C.
• Is Part Of: Electronics letters. Volume 55:Issue 1(2019)
• Page Start: 24
• Page End: 26
• Publication Date: 2019-01-01
• Subjects: image colour analysis -- motion estimation -- rendering (computer graphics) -- computational geometry -- image resolution; scene flow algorithms -- scene flow testing -- scene flow estimation -- colour images -- graphical space -- colour information -- surface deformation -- barycentric coefficients -- ground truth scene flow -- flow vector -- flow maps -- z‐buffer algorithm -- scene flow map -- depth images -- dense scene -- evaluation data generation -- 3D geometry -- natural motion generation -- rigid motion generation -- nonrigid motion generation; Electronics -- Periodicals; 621.381
• DOI: 10.1049/el.2018.6856
• Languages: English
• ISSNs: 0013-5194

Abstract

Abstract : A method which simply generates evaluation data for scene flow testing is presented, in which most algorithms for scene flow estimation require a pair of depth and colour images. We scanned the 3D geometry of a real object and rendered it in graphical space to acquire its depth and colour information. Next, surface deformation to generate natural motion for the scene flow is employed. The surface deformation enabled the generation of non‐rigid as well as rigid motion. Next, barycentric coefficients were exploited to compute the ground truth scene flow. The coefficients enabled the acquired dense scene to flow accurately without dense point tracking and colour information. On the other hand, each flow vector was computed per pixel and a pixel could come from more than one point due to occlusion. We compute the flow maps dividing the geometry into the front area and the occluded area by modifying a z‐buffer algorithm. Experimental results showed that the proposed method reliably provided evaluation data (a pair of depth and colour images and a scene flow map) based on a scanned model with occlusion.