||Abstract Texture registration is an important piece in the pipeline of 3D geometry and texture reconstruction. Its quality is characterized numerically by the covariance matrix of the registration parameters, or by geometric measures, such as the reprojection error. Evaluation with realistic data necessitates datasets with accurately estimated ground truth, difficult to obtain. We propose a new evaluation measure in which the ground truth is represented by the epipolar geometry of texture image pairs which is easy to extract using camera calibration or weak calibration algorithms. We discuss three distance measures based on epipolar geometry and compare them against the reprojection error in an experiment with known (synthetic) ground truth. Using the proposed framework, we also evaluated a texture registration algorithm based on mutual information using the Bouguet camera calibration toolbox for computing the epipolar geometry. We found that the accuracy of the texture registration was less than half-pixel, and we could not detect any bias in the objective functions based on mutual information.