This paper evaluates the insensitivity of different matching costs for stereo correspondence methods to radiometric variations in input images. The authors consider both pixel-based and window-based variants and measure their performance under various conditions, including global and local intensity changes, and noise. Using existing stereo datasets with ground-truth disparities and six new datasets with controlled changes in exposure and lighting, they compare six matching costs (absolute differences, Laplacian of Gaussian, rank filter, mean filter, hierarchical mutual information, and normalized cross-correlation) with three stereo methods (local correlation, semi-global matching, and global graph cuts). The results show that the Rank transform is the best cost for correlation-based methods, hierarchical mutual information performs best for pixel-based global methods, and Rank and Laplacian of Gaussian perform better than hierarchical mutual information for semi-global and global methods in the presence of local radiometric variations. Qualitative analysis of disparity maps confirms these findings, and runtime comparisons highlight the efficiency of the Rank and hierarchical mutual information methods. The study concludes by suggesting future research directions, such as combining the strengths of different costs to handle local radiometric transformations while maintaining sharp depth discontinuities.This paper evaluates the insensitivity of different matching costs for stereo correspondence methods to radiometric variations in input images. The authors consider both pixel-based and window-based variants and measure their performance under various conditions, including global and local intensity changes, and noise. Using existing stereo datasets with ground-truth disparities and six new datasets with controlled changes in exposure and lighting, they compare six matching costs (absolute differences, Laplacian of Gaussian, rank filter, mean filter, hierarchical mutual information, and normalized cross-correlation) with three stereo methods (local correlation, semi-global matching, and global graph cuts). The results show that the Rank transform is the best cost for correlation-based methods, hierarchical mutual information performs best for pixel-based global methods, and Rank and Laplacian of Gaussian perform better than hierarchical mutual information for semi-global and global methods in the presence of local radiometric variations. Qualitative analysis of disparity maps confirms these findings, and runtime comparisons highlight the efficiency of the Rank and hierarchical mutual information methods. The study concludes by suggesting future research directions, such as combining the strengths of different costs to handle local radiometric transformations while maintaining sharp depth discontinuities.