In this paper, we address the problem of increasing the precision of dense direct stereo visual odometry methods. Dense methods need a dense depth map to generate warped images (virtual views) that will match with reference images if the estimated pose is good. Previous works have shown that generating the depth map by machine learning methods leads to very good odometry results. However, machine learning methods generate hallucinated depths even in areas where it is impossible to estimate the depth due to several reasons, like occlusions, homogeneous areas, etc. Generally, this produces wrong depth estimation that leads to errors in odometry estimation. To avoid this problem, we propose a new approach to generate multiple masks that will be combined to discard wrong pixels and therefore increase the accuracy of visual odometry. Our key contribution is to use the multiple masks not only in the odometry computation but also to improve the learning of the neural network for depth map generation. Experiments on several datasets show that masked dense direct stereo visual odometry provides much more accurate results than previous approaches in the literature.