Implementation of symmetric dynamic programming stereo matching algorithm using CUDA

Abstract Stereo correspondence is a computationally intensive procedure, real-time depth map generation for high resolution video is beyond the capability of mainstream CPUs available today. Similar to many other vision algorithms, there is a high degree of parallelism available in most of the correspondence algorithms, making multiprocessor architectures as an obvious choice. Here we describe the implementation of Symmetric dynamic Programming Stereo (SDPS) for graphic processing units (GPU). We discuss the performance results of it on different GPUs. Our implementation focuses on real-time depth map generation for high resolution video pairs, including their lens distortion removal and stereo rectification on GPU.