Narrow Baseline Glsl Multiview Stereo

Our ultimate goal is a system capable of on-line 3D reconstruction from a monocular video and running on commodity hardware. One intrinsic module in such a system is a fast stereo algorithm, which produces depth maps from given views with known camera calibration and pose. Plane-sweep stereo algorithms are well-suited for real-time GPU implementation and can easily be generalised to true multi-image matching, which makes them attractive. This paper presents a simple multiview stereo algorithm implemented for GPUs: GPU-MVS. The details of realising the algorithm with standard (and extended) OpenGL with GL Shading Language (GLSL) are given along with hints on how to make the implementation co-operate well with other, CPU-critical, program modules. The stereo implementation is evaluated on the Middlebury Stereo Evaluation framework for two-view case. The effects of adding more views are investigated, and the final experiment shows some reconstruction results from a complete from-camera-to-3D reconstruction pipeline. To facilitate fair comparisons we will publish the GPU-MVS source code in the Web.