Computer Vision News - October 2019

GSLAM 5 When a new keyframe arrives, a tracking of feature points from its previous frames is applied. If more than 30% of the feature can be tracked, the frame is added to the window, otherwise, the window is sent to pose graph optimization (described below) and a new local window is created to cover the new keyframe.order to better determine the sparse and the smooth regions, the authors incorporate semantic scene information from a DNN- based semantic segmentation. We next review the method in detail. Visual Odometry by Rank-1 Factorization When enough keyframes are added to the window, a global visual odom- etry pipeline is performed. First, the rotations ( R ij ) between every pair of frames in the window are computed. Next, using these rotations, it is pos- sible to retrieve the relative translation and a 3D points from the mapping generated on the previews keyframes. By first assuming a unit depth, the point in the j’th image ( P jk ) has the form: P jk =R ij p k c j =a jk t ij =p k -b jk p jk So, without noise, it needs to hold that Here, c j is the j'th camera position a jk , b jk are scalars, and t ij is the relative translation between the I'th and j'th frame. These two equat ons define two unknows a jk and b jk which are the distance between the I'th and j'th camera and between the point and the j'th camera. To overcome the noise effect, the camera center c j can be approximated as the midpoint of the line seg- ment a-b as shown in the figure below.