We measured the performance of several area-based stereo matching algorithms with noise added to synthetic images. Dense disparity maps were computed and compared with the ground truth using three metrics: the fraction of correctly computed disparities, the mean and standard deviation of the distribution of disparity errors. For a noise-free image, Birchfield and Tomasi's Pixel-toPixel -a dynamic algorithm -performed slightly better than a simple sum-of-absolute differences algorithm (67% correct matches vs 65%) -considered to be within experimental error. A Census algorithm performed worst at only 54%. The dynamic algorithm performed well until the S/N ratio reached 36dB after which its performance started to drop. However, with correctly chosen parameters, it was superior to correlation and Census algorithms until the images became very noisy (∼ 15dB). The dynamic algorithm also ran faster than the fastest correlation algorithms using an optimum window radius of 4 and more than 10 times faster than the Census algorithm.
MotivationThis work was originally motivated by an attempt to implement a stereo algorithm in hardware. Zabih and Woodfill claim that their Census transform is suitable for this [10], but, noting that there are several major groups of stereo matching algorithms and many variants of the individual algorithms within those groups, we felt that a feasibility study to determine which algorithm(s) perform best was needed. We soon discovered, that there has been little serious effort to compare algorithms and provide benchmarks for assessing new algorithms or cost-performance trade-off data which can guide, for example, hardware implementation efforts 1 . Scharstein and Szeliski[9] provide the first thorough 1 The high degree of parallelism present in the matching algorithm makes stereo matching a classic problem for specialized hardware[7].
Figure 1. Disparity maps produced by the Census (left) and Pixel-to-Pixel (right) algorithms against the ground truth (centre).study in this area -comparing over 20 algorithms or variants on two quality measures. Our work extends theirs by including assessments of robustness to noise.When one compares the computed disparity maps in figure 1, it is not at all obvious which algorithm is best. The two disparity maps have about the same number of correct disparities (62.7% and 62.6%) but the standard deviation indicates that the spread of disparity errors for Pixel-to-Pixel is less than for Census (1.5 vs 3.4). Thus metrics that enable quantitative comparisons are important.
The reconstruction of 3D face models is mostly achieved by using 2D images. In this paper, we compare the strengths and weaknesses of different image processing techniques for 3D face generation. It is anticipated that the optimal solution will he applied in the future for 3D face analysis and synthesis. This paper presents binocular stereo using stereo correspondence algorithm or manual triangulation, orthogonal views, and photometric stereo as approaches to 3D face modelling.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.