Recursive Edge-Aware Filters for Stereo Matching

In this study, taxonomy of recursive edge-aware filters (REAF) is provided, with the introduction of new approaches to the state-of-the-art. The one tap recursive filters are classified according to recursion rate calculation, recursion type and the unification of reverse directions. In that manner, eight types of edge-aware recursive filters are defined, where only three of them are addressed in literature so far. Comprehensive analyses are provided based on computational complexity and filter characteristics which affect the use of such filters for various applications. In order to compare the capabilities of these filters, stereo matching, as one of the most common application area of edge-aware filters, is considered and extensive experiments are provided through well known datasets. The evaluation is conducted through large number of stereo pairs on both CPU and GPU with independent parameter optimization of each filter that provides fair comparison. According to the experimental results, advantages of un-normalized recursion for matching accuracy and sequential integration of reverse directions for execution speed are illustrated as important conclusions for future directions of REAFs.