Abstract—3D Video and related technologies like view synthesis, 2D-3D video conversions rely heavily on depth/disparity maps extracted from stereo video content. Innovative Segment-based depth map extraction chain from stereo video content was proposed in [1] giving good trade-off between quality (exactness to the ground truth) and computational complexity. We accelerated this work further by ~150%, both at algorithmic level and by using GP-GPU system (General Purpose processor – Graphics Processing Unit) which is programmed using OpenCL. Our work also compares speedups in terms of cycles consumed GP vs. GPU which goes on to show how GPU can also be utilized for general computations, hence aiding acceleration. Keywords— Depth Estimation, 3D video, OpenCL, GP-GPU, parallel computing, image processing. I. INTRODUCTION O achieve real time performances from complex algorithms it has become necessary to use massively parallel GPUs(Graphics Processing Units) along with GPs (General Processors). This GP-GPU (heterogeneous) computing model is employed in the PC world for graphics, gaming, rendering, server market etc. and now for the handheld/embedded world. To program such systems, Open Computing Language OpenCL [2] an open, royalty free standard for parallel programming of modern processors has been developed. It greatly improves speed and responsiveness for a wide spectrum of applications in numerous market categories from gaming and entertainment to scientific and medical software. We have tried to accelerate our Depth Estimation (DE) chain both by using OCL (OpenCL) on our GP-GPU system and by modifying algorithms which are more suitable to exploit parallelism offered by GPUs. Depth is a complex building block for S3D (stereo 3D) video, the problem of DE in context of S3D video simply stated is to find depth from a given pair of stereo images. This paper is organized as follows, we’ll briefly describe the Graph based DE chain in section II, followed by an introduction to OpenCL in section III. In section IV we G. Visentini is with the Advanced System Technology, STMicroelectronics Agrate, Italy (e-mail: giovanni.visentini@ st.com). A. Gupta is with the Advanced System Technology, STMicroelectronics Greater Noida, India (e-mail: amit.gupta@st.com). describe acceleration achieved using OCL and speedup in terms of cycles consumed by algorithms if they ran on GP compared to GPU. Finally we analyze results and report our conclusions in section V. Section VI gives the direction for future work. II. GRAPH BASED DEPTH ESTIMATION CHAIN The key blocks of our DE chain as described in [1] are shown and described below Fig. 1 Segment (Graph) based Depth estimation chain – double shaded blocks have been accelerated using OpenCL A. Acquisition – Rectification and Color-space conversions A typical stereo acquisition system consists of 2 cameras capturing the same scene, if the intrinsic and extrinsic parameters of this camera setup are known then, based on epipolar geometry corresponding points of the stereo pair can be aligned on the same horizontal line[3]. Through rectification stereo-view pair can be considered as obtained with a parallel camera arrangement and the stereo matching step can be performed on just one dimension rather than many. LEFT RIGHT LEFT RIGHT FEATURES ENHANCEMENT FILTERING EPIPOLAR INVERSE RECTIFICATION SEGMENT-BASED STEREO MATCHING REFINEMENT AND DISPARITY - DEPTH CONVERSION COLOR-BASED SEGMENTATION EPIPOLAR RECTIFICATION Depth Estimation using Open Compute Language (OpenCL) Giovanni Visentini, Amit Gupta T International Conference on Latest Computational Technologies (ICLCT'2012) March 17-18, 2012 Bangkok 55