A method based on multiple adaptive windows to improve the determination of disparity maps

The correspondence problem is one of main
topics in stereo vision that, despites being studied for many
years, is still in progress.
In this paper, we present a new method that computes
the disparity map. Our method is local (i.e. only information
gathered in the close neighborhood is used) and is
based on image statistics. More specifically it combines multiple
adaptive windows and local statistical measures in order
to optimize the quality of the computation of disparity
maps. As the size of windows is critical for local methods,
we propose an algorithm that modifies the size of windows
surrounding the pixel of interest to capture enough information
in regions with low texture energy. This process is based
on statistical measures (mean and horizontal/vertical average
deviation) taken in the windows of the original stereo
images.
In the last section we show that our method performs very
well compared to other existing local methods.