Architectures for Stereo Vision - Signal Processing Systems - page 473

Fig. 1 Results for the stereo correspondence problem: ( a ) left rectified input image (raw input
images taken from [ 20 ] ), ( b ) disparity map after left/right check where white denotes disparities
marked as invalid, ( c ) false color representation of the disparity map, ( d ) untextured 3D view
generated from the disparity map of ( b )
difference ( the disparity) in conjunction with a known stereo camera calibration
allows to infer the depth information. Figure 1 gives an example.
The importance of stereo matching has been underlined by Szeliski and
Scharstein stating that it is “one of the most widely studied and fundamental
problems of computer vision” [ 83 ] . Active research in this field has resulted in a
wide range of disparity estimation algorithms using radically different approaches.
Recently, a general taxonomy has been introduced [ 81 ] including a comprehensive
survey, what resulted in the on-going online Middlebury benchmark [ 80 ] . Further
surveys evaluated different algorithms and variations thereof [ 39 , 86 ] . Major focus
was the quality of the stereo matching in terms of accuracy, density of the disparity
map, and robustness.
However, advances in robustness and accuracy were accompanied with signif-
icant increases in complexity and computational requirements making the use of
specialized implementations for many of today's real-time applications an absolute
necessity. Surveys on efficient implementations for selected types of algorithms
have been conducted [ 29 , 59 , 66 , 86 ] and many more specialized implementations
and architectures for individual algorithms and applications have been proposed.