A Multifeature Correspondence Algorithm Using Dynamic Programming
C. V. Jawahar, P. J. Narayanan · 2002
Correspondence between pixels is an important problem in stereo vision. Several algorithms have been proposed to carry out this task in literature. Almost all of them employ only gray-values. We show here that addition of primary or secondary evidence maps can improve the correspondence computation. However any particular combination is not guaranteed to provide proper results in a general sitiua- tion. What one needs is a mechanism to select the evidences which are apropriate for a particular pair of images. We present an algorithm for stereo correspondence that can take advantage of different image features adaptively for matching. A match measure combining different individual measures computedfrom different features is used by our al- gorithm. The advantages of each feature can be combined in a single correspondence computation. We describe an unsupervised scheme to compute the relevance of each fea- ture to a particular situation, given a set of possibly useful features. We present an implementationof the scheme using dynamic programming for pixel-to-pixel correspondence. views. Area based matching algorithms are good for scenes with good texture, edge based algorithms are good when edges are present, etc. There have been some attempts to formulate the correspondence problem in a general statisti- cal framework using the maximum likelihood estimates for the pixels (4) or by estimating the Bayesian priors from the intensity distribution (2). These, in essence, compute a sim- ple similarity measure between gray-values of pixels in both the images and find the optimal matches by imposing new constraints, using penalty terms for the occluded pixels. We present a scheme in this paper to adaptively select the combination of features that work best for a specific pair of images. The selection starts with a superset of fea- tures that could be relevant for matching between the two images. These could include intensity along multiple spec- trums such as different colour bands, edge strength, tex- ture measures, etc. The matching measures computed from these diverse features are combined, with appropriate im- portances assigned to each in the form of a weight, to yield a single measure of similarity or dissimilarity between two candidate pairs of pixels. The weight of a particular fea- ture encodes its relevance or importance in matching the pair of images. We also present a scheme for estimating the weights for the features used which converges fast on typical images. The importance of integrating multiple feature measures for stereo correspondence has been recognized in the litera- ture (4, 5, 6), but practical implementations involving mul- tiple features are rare. We introduced the framework of gen- eralised correlation to combine diverse types of features in a fle xible manner (5). It was quite successful in combin- ing multiple features under a correlation framework. The importances of the individual feature measures were, how- ever, hand computed with no fle xibility to adapt to a pair of images automatically. We later devised a technique to esti- mate the importances of each feature based on the siutation under the correlation framework (6). An adaptive, non-supervised scheme to estimate the rel- evances of the feature measures used depending on the re- sults of the matching is also presented in this paper. We present the results from implementing our scheme using dynamic programming. The methodology differs consider-