Multiwavelets in the Context of Hierarchical Stereo Correspondence Matching Techniques

This paper presents an evaluation of different types and families of multiwavelets in stereo correspondence matching. First, the paper introduces two hierarchical stereo matching techniques based on balanced and respectively unbalanced multiwavelet transforms, which employ normalized cross correlation to search for disparities. Different multiwavelet families, with different properties and filter types are evaluated, such as balanced versus unbalanced multiwavelets and symmetric-symmetric versus symmetric-antisymmetric multiwavelets. Each approximation subband carries a different spectral content of the original image and the information in the basebands of the multiwavelet transform is less sensitive to the shift variability of the multiwavelet transform. This can be exploited in order to improve the accuracy of the initial disparity map. As this initial disparity map is estimated at the lowest resolution, it needs to be progressively propagated to higher resolution levels. As a result, the search at high resolution levels is significantly reduced, thereby reducing the computational cost of the overall process and improving the reliability of the final disparity map. The evaluation of different types and families of multiwavelets shows that unbalanced multiwavelets produce a smoother disparity map with less mismatch errors compared to balanced multiwavelets. Finally, the paper introduces a third technique, which replaces normalized cross correlation with a better performing global error energy minimization algorithm operating based on a similar hierarchical technique. The results show that the multiwavelet techniques produce a smoother disparity map with less mismatch errors compared to applying a similar matching algorithm in either the spatial and/or the wavelet domains. The performance of the proposed algorithms is also compared against several state-of-theart techniques from the Middlebury database. KeywordsMultiwavelets, Correspondence matching, Disparity estimation, Stereo vision.