Computer Vision News - October 2019

Research 6 This boiled down to the linear equation of C j =0.5*(a jk t ij +p k -b jk p jk ) By solving the linear equation (for a jk , b jk ) all the quantities above are known. Now we can relax the assumption of unit depth, and by performing several manipulations the authors formalize a rank one factorization where the right-hand side is factorized into a multiplication of the columns of the camera centers (of each camera) with a row of the depth values of each of the points in the scene. This gives a global solution to the depth and centers which will be refined later. Pose Graph Optimization The final stage of the method, similarly to the loop closure thread in other methods, is to generate a globally consistent map using the local maps generated from previous steps. To this end, the author suggests building a pose graph. This graph contains vertices which are the keyframes and edges, that is the 3D similarity between two keyframe's local map. This graph is generated using several heuristics that define the existence of a relation between every two nodes in the graph. After building such graph, the authors use a global pose graph optimization to determine the global rotation between the edges, the scale, and the translation. To be able to be relevant for SLAM applications, the optimization is done using L-1 norm minimization by linear programming. The method can be summarized by the following figure: