The paper introduces a novel approach to the correspondence problem in computer vision, which is fundamental for tasks like stereo depth computation and optical flow algorithms. The authors propose using non-parametric local transforms as the basis for correlation, which rely on the relative ordering of local intensity values rather than the intensity values themselves. This approach can tolerate a significant number of outliers, leading to improved performance near object boundaries compared to conventional methods like normalized correlation. Two non-parametric local transforms are introduced: the *rank transform* and the *census transform*. The rank transform measures local intensity by counting the number of pixels in a neighborhood with lower intensity than the center pixel. The census transform summarizes local image structure by mapping the neighborhood to a bit string representing pixels with lower intensity than the center pixel. Both transforms are designed to handle factionalism, where a minority of pixels have a different intensity distribution from the majority, by focusing on the set of pixel comparisons rather than the intensity values themselves. The paper demonstrates the utility of these transforms through empirical results on synthetic and real data. It shows that the non-parametric local transforms outperform normalized correlation in terms of accuracy near object boundaries. The authors also discuss related work and plan to extend their approach by combining multiple non-parametric transforms and exploring higher-order differences and total ordering of pixel intensities.The paper introduces a novel approach to the correspondence problem in computer vision, which is fundamental for tasks like stereo depth computation and optical flow algorithms. The authors propose using non-parametric local transforms as the basis for correlation, which rely on the relative ordering of local intensity values rather than the intensity values themselves. This approach can tolerate a significant number of outliers, leading to improved performance near object boundaries compared to conventional methods like normalized correlation. Two non-parametric local transforms are introduced: the *rank transform* and the *census transform*. The rank transform measures local intensity by counting the number of pixels in a neighborhood with lower intensity than the center pixel. The census transform summarizes local image structure by mapping the neighborhood to a bit string representing pixels with lower intensity than the center pixel. Both transforms are designed to handle factionalism, where a minority of pixels have a different intensity distribution from the majority, by focusing on the set of pixel comparisons rather than the intensity values themselves. The paper demonstrates the utility of these transforms through empirical results on synthetic and real data. It shows that the non-parametric local transforms outperform normalized correlation in terms of accuracy near object boundaries. The authors also discuss related work and plan to extend their approach by combining multiple non-parametric transforms and exploring higher-order differences and total ordering of pixel intensities.