Computer Vision News - February 2019

Visual SLAM approaches Deep Learning is in the process of taking over the lead in the geometric vision field. We will now examine how Deep Learning methods and their specific capabilities can be used to replace individual elements / stages of the Visual SLAM pipeline, and consider the possibility of combining those solutions into an overall Deep Learning Visual SLAM. Localization and Depth Estimation: Localization and depth estimation are crucial processes for Automated Driving. Three approaches exist: Supervised, Unsupervised and Semi-supervised. Supervised is the most common approach: This approach estimates depth using methods very similar to those of semantic segmentation, inspired by classification networks. The loss function is usually a continuous regression function, but some attempts have treated depth learning as a form of classification problem, giving different degrees of discreteness to depth learning. The unsupervised approach relies on running a projection function between different displays, for instance, by estimating movement between different views. Semi-i-supervised approaches combine elements of both. The table below summarizes the results of three methods, each implementing a different approach, on the KITTI dataset. 7 Research Computer Vision News Visual SLAM for Automated Driving