This paper describes an efficient method for dense matching of two or three 
images. After some investigations in different similarity measures we propose a 
modification of Semi-Global Matching, which uses a simple energy function that 
implies piecewise smoothness but no ordering and gives promising results in 
practice. Our improvements include a symmetric and hierarchical matching 
strategy and allow an efficient generalization of the stereo matching problem to 
trinocular surface reconstruction. Finally, we present results for synthetic and 
for real images.