Stereo Matching with Non � Linear Di usion
نویسندگان
چکیده
One of the central problems in stereo matching and other image registration tasks is the selection of opti mal window sizes for comparing image regions This paper addresses this problem with some novel algo rithms based on iteratively di using support at dif ferent disparity hypotheses and locally controlling the amount of di usion based on the current quality of the disparity estimate It also develops a novel Bayesian estimation technique which signi cantly outperforms techniques based on area based matching SSD and regular di usion We provide experimental results on both synthetic and real stereo image pairs Introduction and related work Most area based approaches to the stereo corre spondence problem perform the following three tasks For each disparity under consideration compute a per pixel matching cost Aggregate support spatially e g by summing over a window or by di usion Across all disparities nd the best match based on the aggregated support The focus of this paper is the second step aggrega tion of support A central problem is to nd the op timal size of the support region If the region is too small a wrong match might be found due to ambigui ties and noise If the region is too big it can no longer be matched as a whole due to foreshortening and oc clusion with the result of lost detail and blurring of object boundaries Kanade and Okutomi have proposed adaptive windows square windows that extend by di erent amounts in each of four directions The optimal win dow size is found by a greedy algorithm gradient de scent based on an estimate of disparity uncertainty in the current window In this paper we propose a Supported by NSF grant IRI di erent approach aggregating support with a non uniform di usion process A support region can either be two dimensional at a xed disparity favoring fronto parallel surfaces or three dimensional in x y d space supporting slanted surfaces Two dimensional evidence aggregation has been done using square windows traditional Gaus sian convolution and windows with adaptive sizes Three dimensional support functions that have been proposed include limited disparity di erence limited disparity gradient and Prazdny s coher ence principle which can be implemented using two di usion processes Some matching costs such as correlation and non parametric measures are de ned over a certain area of support and thus combine the cost and aggregation steps into one Measures that can be accumulated in a separate step however have the following advantages E ciency the measure can be aggregated with a single convolution or box lter operation Parallelizability the aggregation step can be im plemented on highly parallel architectures using local iterative di usion Adaptability the measure can be aggregated over locally di erent support regions using either ad justable size windows or a non uniform di u sion process this paper Other stereo techniques include hybrid and itera tive techniques such as stochastic search as well as hierarchical methods More than two images are used in multiframe stereo to increase stability of the algorithm A long version of this paper containing more references is available as Cornell CS TR For a general survey of stereo vision methods see Disparity space and SSD Support for a match is de ned over a three dimensional disparity space E x y d Formally we
منابع مشابه
Stereo Matching with Non - Linear Di usion DANIEL
One of the central problems in stereo matching (and other image registration tasks) is the selection of optimal window sizes for comparing image regions. This paper addresses this problem with some novel algorithms based on iteratively di using support at di erent disparity hypotheses, and locally controlling the amount of di usion based on the current quality of the disparity estimate. It also...
متن کاملLocal Scale Controlled Anisotropic Di usion with Local Noise Estimate for Image Smoothing and Edge Detection
A novel local scale controlled piecewise linear diffusion for selective smoothing and edge detection is presented. The di usion stops at the place and time determined by the minimum reliable local scale and a spatial variant, anisotropic local noise estimate. It shows nisotropic, nonlinear di usion equation using di usion coe cients/tensors that continuously depend on the gradient is not necess...
متن کاملAnisotropic Di usion of Noisy Surfaces and Noisy Functions on Surfaces
We present a uni ed anisotropic geometric di usion PDEmodel for smoothing (fairing) out noise both in triangulated 2-manifold surface meshes in IR and functions de ned on these surface meshes, while enhancing curve features on both by careful choice of an anisotropic di usion tensor. We combine the C limit representation of Loop's subdivision for triangular surface meshes and vector functions o...
متن کاملA nonoscillatory numerical scheme based on a general solution of 2-D unsteady advection–di usion equations
A non-oscillatory numerical scheme based on a general solution of unsteady advection–di usion equations is presented. A general solution for initial value problems of linear two-dimensional unsteady advection–di usion equations is obtained using the spectral method. The resulting numerical scheme is explicit with respect to time, and ful lls the Patankar’s positive coe cients condition for any ...
متن کاملOptical di usion of cw, pulse and density waves in scattering media and comparisons with radiative transfer
Transport theory's equation of transfer is a well-accepted model for wave scattering in a random medium. However, exact solutions are only available for a small class of special cases (isotropic scattering in in nite and semiin nite domains) and thus, analytic solutions for a wide variety of geometries and range of parameters which correspond to more physically relevant situations still remain ...
متن کامل