This paper proposes a new approach to the correspondence problem in computer vision using non-parametric local transforms. The correspondence problem involves finding matching pixels between two images of the same scene. Traditional methods, such as normalized correlation, often struggle with object boundaries due to the presence of outliers. Non-parametric local transforms, which rely on the relative ordering of local intensity values rather than their actual values, are more robust to outliers and can provide better performance near object boundaries. The paper introduces two non-parametric local transforms: 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 spatial structure by creating a bit string that represents the relative intensities of neighboring pixels. These transforms are invariant to changes in image gain and bias, making them suitable for stereo vision applications. The paper demonstrates the effectiveness of these transforms on both synthetic and real data. Empirical results show that the non-parametric transforms outperform normalized correlation in terms of accuracy and robustness, especially near object boundaries. The transforms are also shown to be effective in computing stereo depth from real images, with results displayed in figures showing depth maps generated using the rank and census transforms. The paper also discusses related work and planned extensions, including the combination of multiple non-parametric transforms into a vector of measures and the use of higher-order differences to improve the transforms. Efficient algorithms for implementing these transforms are also explored. The results show that the non-parametric local transforms can compute stereo depth with high accuracy and efficiency, making them a promising approach for computer vision applications.This paper proposes a new approach to the correspondence problem in computer vision using non-parametric local transforms. The correspondence problem involves finding matching pixels between two images of the same scene. Traditional methods, such as normalized correlation, often struggle with object boundaries due to the presence of outliers. Non-parametric local transforms, which rely on the relative ordering of local intensity values rather than their actual values, are more robust to outliers and can provide better performance near object boundaries. The paper introduces two non-parametric local transforms: 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 spatial structure by creating a bit string that represents the relative intensities of neighboring pixels. These transforms are invariant to changes in image gain and bias, making them suitable for stereo vision applications. The paper demonstrates the effectiveness of these transforms on both synthetic and real data. Empirical results show that the non-parametric transforms outperform normalized correlation in terms of accuracy and robustness, especially near object boundaries. The transforms are also shown to be effective in computing stereo depth from real images, with results displayed in figures showing depth maps generated using the rank and census transforms. The paper also discusses related work and planned extensions, including the combination of multiple non-parametric transforms into a vector of measures and the use of higher-order differences to improve the transforms. Efficient algorithms for implementing these transforms are also explored. The results show that the non-parametric local transforms can compute stereo depth with high accuracy and efficiency, making them a promising approach for computer vision applications.