PROJECT TITLE :

Deep Visual Odometry with Adaptive Memory

ABSTRACT:

A novel deep visual odometry (VO) method that takes into account global information by selecting memory and refining poses is presented here. The currently available learning-based methods approach the VO task as if it were a pure tracking problem, recovering camera poses from image snippets. This approach results in a significant accumulation of error. The availability of global information is absolutely necessary in order to reduce accumulated errors. On the other hand, it can be difficult to effectively preserve such information across systems from beginning to end. In order to overcome this obstacle, we have developed an adaptive memory module. This module saves information in a neural memory analogy in a progressive and adaptive manner, moving it from a local to a global level. This gives our system the ability to process long-term dependencies. An additional refining module uses the global information stored in the memory to further improve upon the accuracy of the previously obtained results. In order to select the appropriate features for each view based on their co-visibility in the feature domain, we make use of a spatial-temporal attention, which is guided by the results of earlier analyses. Specifically, our architecture, which consists of modules for tracking, remembering, and refining, works beyond tracking in its capabilities. Experiments conducted on the KITTI and TUM-RGBD datasets show that our method outperforms methods considered to be state-of-the-art by significant margins and produces results that are competitive with those produced by traditional methods when applied to regular scenes. In addition, our model is able to achieve outstanding performance in difficult scenarios such as texture-less regions and abrupt motions, both of which are areas in which traditional algorithms have a tendency to fail.