Deep Visual Odometry with Adaptive Memory

TL;DR

This paper introduces a deep visual odometry method that uses an adaptive memory module to incorporate global information, significantly reducing error accumulation and improving performance in challenging scenarios.

Contribution

The novel adaptive memory module enables end-to-end deep VO systems to effectively preserve and utilize global information for long-term dependency modeling.

Findings

Outperforms state-of-the-art methods on KITTI and TUM-RGBD datasets.

Achieves robust performance in texture-less regions and abrupt motions.

Produces competitive results against classic VO approaches.

Abstract

We propose a novel deep visual odometry (VO) method that considers global information by selecting memory and refining poses. Existing learning-based methods take the VO task as a pure tracking problem via recovering camera poses from image snippets, leading to severe error accumulation. Global information is crucial for alleviating accumulated errors. However, it is challenging to effectively preserve such information for end-to-end systems. To deal with this challenge, we design an adaptive memory module, which progressively and adaptively saves the information from local to global in a neural analogue of memory, enabling our system to process long-term dependency. Benefiting from global information in the memory, previous results are further refined by an additional refining module. With the guidance of previous outputs, we adopt a spatial-temporal attention to select features for each view based on the co-visibility in feature domain. Specifically, our architecture consisting of Tracking, Remembering and Refining modules works beyond tracking. Experiments on the KITTI and TUM-RGBD datasets demonstrate that our approach outperforms state-of-the-art methods by large margins and produces competitive results against classic approaches in regular scenes. Moreover, our model achieves outstanding performance in challenging scenarios such as texture-less regions and abrupt motions, where classic algorithms tend to fail.