Although deep learning methods have been widely applied in slam visual odometry (VO) over the past decade with impressive improvements, the accuracy remains limited in complex dynamic environments. In this paper, a composite mask-based generative adversarial network (CMGAN) is introduced to predict camera motion and binocular depth maps. Specifically, a perceptual generator is constructed to obtain the corresponding parallax map and optical flow between two neighboring frames. Then, an iterative pose improvement strategy is proposed to improve the accuracy of pose estimation. Finally, a composite mask is embedded in the discriminator to sense structural deformation in the synthesized virtual image, thereby increasing the overall structural constraints of the network model, improving the accuracy of camera pose estimation, and reducing drift issues in the VO. Detailed quantitative and qualitative evaluations on the KITTI dataset show that the proposed framework outperforms existing conventional, supervised learning and unsupervised depth VO methods, providing better results in both pose estimation and depth estimation.