This work has been funded by a grant from NVIDIA.
We have implemented a basic binocular stereo algorithm in CUDA. CUDA is an NVIDIA product which allows programs written in a modified C language to be executed on the GPU. Using CUDA, our stereo algorithm outperforms our previous GPU stereo work, which was implemented using programmable shaders (Kim et al. 2007). Our stereo runs at 40Hz on 640x480 images at 50 disparities, computing left and right disparity maps and performing left-right consistency validation.
A brief video demonstrates our stereo algorithm performing a live real-time reconstruction of a dynamic scene. The video is in the WMV format.
Source code is made available under the BSD license. This code uses the CUDA SDK.
For a slightly different version of our algorithm with a detailed paper describing the implementation, please see "CUDA Stereo Imaging" at http://openvidia.sourceforge.net/index.php/OpenVIDIA.