Computer Vision News - June 2019

Every month, Computer Vision News reviews a research paper from our field. This month we have chosen BA-Net: Dense Bundle Adjustment Network . We are indebted to the authors ( Chengzhou Tang, Ping Tan ) for allowing us to use their images. The paper was presented at ICLR 2019 a few weeks ago and is found here . BA net- Bundle Adjustment Network Structure from motion is one of the fundamental branches in computer vision. Surprisingly, classical solutions still outperform deep learning solutions in many structure from motion tasks. One and very important example of such task is bundle adjustment (BA) , which is a highly non-convex optimization. The challenges in BA are the sensitivity to initial correspondences, non-linearity of the objective, sensitivity to outliers and more. In the BA-net paper, which was presented at ICLR 2019 , the authors suggest a novel deep learning architecture that aims to solve the bundle adjustment problem. It uses BA as a differentiable layer in the network. The main novelty in the paper is to learn a feed-forward network to predict the damping factor of the Levenberg-Marquardt (LM) algorithm which makes the whole pipeline differentiable. Bundle adjustment To better understand the novelty of the paper, we first explain the conventional bundle adjustment. Generally, almost every structure from motion algorithm uses bundle adjustment to refine its final solution. Given a set of N images I i , initial guesses for N camera matrices T i and an initial guess for a set of K 3-D points p j . The geometric bundle adjustment minimizes the sum of reprojection errors of the form: Where q i,j is the 2-D point in the i'th image, corresponding to the 3-D point p j . ( ,∙) denotes the projection operator of a 3-D point into the image plane, using camera matrix T. The optimization variables are the camera matrices T 1 ,..,T N and the 3-D points p 1 ,..,p K . This optimization is usually called geometric BA . On the other hand, sensitivity to outliers and texture has motivated the emergence of photometric bundle adjustment . In contrast to geometric, the photometric BA relies on maximizing the photometric (pixel intensity value) 4 Research by Amnon Geifman Computer Vision News Research =1 | , − , |