Parallel Algorithm Using Opencl for Depth Estimation of Image

Abstract

Depth maps have a tremendous importance in Robot Navigation, Real-time tracking of scenes, 2D-3D con- versions needed for gaming as well as in animated movies and much more. A fairly good implementation of Segment based Depth Estimation chain using GPU(Graphic Processing Unit) computing has been presented in the work by G. Visentini and A. Gupta (2012), but it lacks in the segmentation step to be implemented in parallel and is a bit naive in other steps like bilateral filtering and Stereo Matching in terms of harnessing the GPU power optimally. We present a novel parallel implementation of graph based segmentation using the Boruvka's Minimum Spanning tree(MST) algorithm in Open Compute Language(OpenCL). We have optimized the GPU performance by making a significant use of shared or local memory of the GPU. We choose OpenCL for parallel computing rather than CUDA due to it's portability on heterogeneous systems (multi-core CPU-GPU integrated) and it's open source environment which allows it to be used on GPUs and CPUs of any brand. The Open- CL program can also be run on handheld embedded systems and FPGAs (Altera has recently launched it's OpenCL SDK for FPGAs). OpenCL also has inbuilt vector data-types, the use of which makes the processing the R,G and B values of an image