Visual odometry with depth-wise separable convolution and quaternion neural networks

Monocular visual odometry is a fundamental problem in computer vision and it was extensively studied in literature. The vast majority of visual odometry algorithms are based on a standard pipeline consisting in feature detection, feature matching, motion estimation and local optimization. Only recently, deep learning approaches have shown cutting-edge performance, replacing the standard pipeline with an end-to-end solution. One of the main advantages of deep learning approaches over the standard methods is the reduced inference time, that is an important requirement for the application of visual odometry in real-time. Less emphasis, however, has been placed on memory requirements and training efficiency. The memory footprint, in particular, is important for real world applications such as robot navigation or autonomous driving, where the devices have limited memory resources. In this paper we tackle both aspects introducing novel architectures based on Depth-Wise Separable Convolutional Neural Network and deep Quaternion Recurrent Convolutional Neural Network. In particular, we obtain equal or better accuracy with respect to the other state-of-the-art methods on the KITTI VO dataset with a reduction of the number of parameters and a speed-up in the inference time.