Learning-based Low Light Image Enhancement for Visual Odometry

For the low light environment, performing good visual odometry still remains an open challenge. The main difficulties in low light situations lie in that the limitations of the sensors and the buried features in the dark, which hamper to achieve successful detecting and tracking of interest points. Although there are some pioneering works proposed on image enhancement for low light situations, there are comparatively little works focusing on visual odometry in low light scenes. Here, we address this problem from a deep learning view, for which we train a deep neural network for image enhancement to obtain the enhanced image sequences for VO. In contrast to the end-to-end training scheme, the network is designed based on the physical Retinex model, and the L₁ distance loss and structure similarity index measure (SSIM) loss are incorporated into network training. We validate the performance of the enhancement by evaluating the enhanced sequences by state-of-the-art VO algorithm, ORB-SLAM.