DeepAVO: Efficient pose refining with feature distilling for deep Visual Odometry

The technology for Visual Odometry (VO) that estimates the position and orientation of moving object through analyzing image sequences captured by on-board cameras, has been well investigated with rising interest in autonomous driving. This paper studies monocular VO from perspective Deep Learning (DL). Unlike most current learning-based methods, our approach, called DeepAVO, is established on intuition features contribute discriminately to different motion patterns. Specifically, we present a novel four-branch network learn rotation translation leveraging Convolutional Neural Networks (CNNs) focus quadrants optical flow input. To enhance ability feature selection, further introduce an effective channel-spatial attention mechanism force each branch explicitly distill related information specific Frame (F2F) estimation. Experiments various datasets involving outdoor driving indoor walking scenarios show proposed DeepAVO outperforms state-of-the-art methods large margin, demonstrating competitive performance stereo algorithm verifying promising potential generalization.