REAL-TIMESTEREO CORRESPONDENCE USINGA TRUNCATED SEPARABLE LAPLACIANKERNEL APPROXIMATIONON GRAPHICSHARDWARE

We present anovel real-time stereo algorithm thatachieves both goodquality results andveryhighdisparity estimation throughput onthegraphics processing unit(GPU). Asthekeyideaofthis paper, atruncated separable approximation toanisotropic Laplacian kernel isproposed. Thistruncated 2DLaplacian kernel variant combines theadvantages oflarge support windows andshiftable windows,while support-weights ongeometric proximity canstill be appropriately applied toeachpixel intruncated support windows. Ourmethodoutperforms previous GPU-based local stereo methods andevensomemethods using global optimization onthebenchmark Middlebury stereo database. Because ofitsseparable andregular property, theproposed kernel canbeveryefficiently implemented onGPUs.Ouroptimized implementation completely running onan Nvidia GeForce 7900graphics cardachieves over668million disparity estimations persecond (Mde/s) including alltheoverhead, about 2.3to13.4 times faster thantheexisting GPU-based solutions.