Computer Vision News - June 2019

consistency and estimates the correspondences implicitly. Given a set of camera matrices T and a set of K points (in I 1 pixel coordinates) q. The photometric error is then defined as: Where d j is the depth of the point q j , hence d j q j upgrades the pixel q j to its 3D coordinates. The photometric bundle adjustment minimizes the sum of the photometric reprojection errors. With this intuition in mind, the authors define a differential pipeline that enables to exploit the power of deep learning. We next explain the method. Understanding the model The authors suggest a slight change to the photometric BA . Instead of minimizing the photometric error, they offer to minimize the difference (error) of features related to a specific pixel, i.e. feature-metric difference of aligned pixel, that is: Where in this error term, everything stays the same as before, except that now F i (q j ) denotes a (learnable) feature pyramid i.e. a feature vector across multiple scales corresponds to the q j pixel in the i'th image. The minimization is over the camera parameters and depths in the sum , , , The above figure best explains the method. The input to the pipeline is a sequence of images. Then, every image is transferred through a DRN-54 network (see architecture in the next figure) to learn a feature pyramid. At the same time, a convolutional network generates multiple basis depth maps Computer Vision News 5 Research Computer Vision News A novel architecture that enables end-to-end differentiable bundle adjustment BA-Net: Dense Bundle Adjustment Network , , , = , − 1 ( , , , = , − 1 (