Dense Stereo Matching

The identification of homologous pixels in a stereo pair is a fundamental process in automated 3D reconstruction from images. In this project, we focus on dense matching using local optimization methods.

Our approach is based on a combination of state-of-the-art algorithms and innovative concepts that enhance both cost computation and its aggregation. The cost is derived from the absolute color difference and a census-type transform applied to the gradient images.

The aggregation is performed using cross-based windows that can be expanded with a dynamic threshold. Optimization is carried out through scanline processing combined with a winner-takes-all technique. Errors and occlusions are corrected using established methods.