Biometrics systems typically work best in settings where probe samples are captured in the same manner as the training set. When biometrics are acquired under different conditions or with different sensors, naïve approaches to recognition perform poorly. Coupled mappings have been introduced for performing face recognition across different resolutions, and learn a common subspace between different domains. In this paper, we introduce Maximum-Margin Coupled Mappings (MMCM), which aims to learn projections such that there is a margin of separation between pairs of cross-domain data from the same class and pairs of cross-domain data from different classes. While coupled mapping techniques have traditionally been used for matching face images at different resolutions, we demonstrate that MMCM is effective for cross-sensor biometric matching as well.