This paper evaluates the insensitivity of different matching costs in stereo matching to radiometric variations in input images. The authors consider both pixel-based and window-based matching costs and assess their performance under global intensity changes, local intensity changes, and noise. They use existing stereo datasets with ground-truth disparities and six new datasets with controlled exposure and lighting changes. The evaluation is performed using three stereo methods: local, semi-global, and global. The paper discusses various matching costs, including absolute differences, filters (LoG, Rank, Mean), mutual information (MI), and normalized cross-correlation (NCC). The study finds that the Rank transform performs best for correlation-based methods, while hierarchical mutual information (HMI) is best for pixel-based global methods like SGM and GC in the presence of global brightness changes and noise. For local radiometric variations, Rank and LoG perform better than HMI. The authors also evaluate the performance of different stereo algorithms, including a local correlation method (Corr), a semi-global matching method (SGM), and a global method using graph cuts (GC). The results show that SGM and GC perform better than Corr in most cases. The study also highlights the importance of robustness to outliers and the impact of radiometric changes on stereo matching. The paper concludes that while some matching costs perform well under certain conditions, none are highly effective at handling strong local radiometric changes caused by changes in light source positions. Future work includes testing other matching costs that can handle radiometric differences, such as the census transform and approximations of mutual information.This paper evaluates the insensitivity of different matching costs in stereo matching to radiometric variations in input images. The authors consider both pixel-based and window-based matching costs and assess their performance under global intensity changes, local intensity changes, and noise. They use existing stereo datasets with ground-truth disparities and six new datasets with controlled exposure and lighting changes. The evaluation is performed using three stereo methods: local, semi-global, and global. The paper discusses various matching costs, including absolute differences, filters (LoG, Rank, Mean), mutual information (MI), and normalized cross-correlation (NCC). The study finds that the Rank transform performs best for correlation-based methods, while hierarchical mutual information (HMI) is best for pixel-based global methods like SGM and GC in the presence of global brightness changes and noise. For local radiometric variations, Rank and LoG perform better than HMI. The authors also evaluate the performance of different stereo algorithms, including a local correlation method (Corr), a semi-global matching method (SGM), and a global method using graph cuts (GC). The results show that SGM and GC perform better than Corr in most cases. The study also highlights the importance of robustness to outliers and the impact of radiometric changes on stereo matching. The paper concludes that while some matching costs perform well under certain conditions, none are highly effective at handling strong local radiometric changes caused by changes in light source positions. Future work includes testing other matching costs that can handle radiometric differences, such as the census transform and approximations of mutual information.