An efficient region-adaptive stereo matching algorithm

This paper presents a simple and efficient region adaptive local stereo matching algorithm which focuses on different cost aggregation strategies for different regions. Cost aggregation strategy for local matching methods generally applies a support window which contains corresponding weights for each pixel in the window. However, these algorithms implement the same technique on all pixels, costing heavy computation on searching the best disparity. In our research, we found that special region should adopt specified technique in order to get high accuracy when doing cost aggregation. So we decide to combine different methods for different areas. Moreover, our method is aimed at realtime requirements on 3D information retrieval in many applications. For real-time purpose, we apply census transform method along depth discontinuities and a fast segmentation-based method for other low texture areas.