Stereo Matching Algorithm Based on a Generalized Bilateral Filter Model

Stereo matching is a kernel problem in stereo vision systems. Stereo algorithms can be roughly classified into local and global approaches. Local algorithms use Winner-Take-All strategy, simply taking disparity level that minimizes the aggregation costs. In this paper we present a local stereo matching algorithm with an adaptive cost aggregation strategy based on a generalized bilateral filter model. The range weight computation in the original bilateral filter is extended by the inner and outer weighted average processes. A pixel is assigned a high range weight to the central pixel not only if the patches of the two pixels are similar but also if the neighbouring patches around the two pixels are similar. The final range weight could more accurately reflect the similarity of considering two pixels. Different cost aggregation methods can be easily derived from the model by modifying parameters. In experimental section we compare four cost aggregation methods based on the generalized model and give conclusions which demonstrate the effectiveness of our proposed strategy.