Over the last decade, one of the most relevant public datasets for evaluating\nodometry accuracy is the KITTI dataset. Beside the quality and rich sensor\nsetup, its success is also due to the online evaluation tool, which enables\nresearchers to benchmark and compare algorithms. The results are evaluated on\nthe test subset solely, without any knowledge about the ground truth, yielding\nunbiased, overfit free and therefore relevant validation for robot localization\nbased on cameras, 3D laser or combination of both. However, as any sensor\nsetup, it requires prior calibration and rectified stereo images are provided,\nintroducing dependence on the default calibration parameters. Given that, a\nnatural question arises if a better set of calibration parameters can be found\nthat would yield higher odometry accuracy. In this paper, we propose a new\napproach for one shot calibration of the KITTI dataset multiple camera setup.\nThe approach yields better calibration parameters, both in the sense of lower\ncalibration reprojection errors and lower visual odometry error. We conducted\nexperiments where we show for three different odometry algorithms, namely\nSOFT2, ORB-SLAM2 and VISO2, that odometry accuracy is significantly improved\nwith the proposed calibration parameters. Moreover, our odometry, SOFT2, in\nconjunction with the proposed calibration method achieved the highest accuracy\non the official KITTI scoreboard with 0.53% translational and 0.0009 deg/m\nrotational error, outperforming even 3D laser-based methods.\n