Computer Vision News - June 2019

As you can see, BA-net outperforms all the other methods . They show a significant advantage in camera rotation and translation estimation (first 3 rows) as well as in the root mean square error of the depth values. You can also see below some qualitative results on the ScanNet dataset that show the recovered depth maps by the paper's method and the DeMoN method. The circles demonstrate that the BA-net method tend to recover more shape details. Moreover, the results seem sharper and it looks like the BA net gives a better estimation for the depth map (at least in this specific example). Additional results are in the paper. Conclusion: The paper presents a novel architecture that enables end-to-end differentiable bundle adjustment . It allows to deal with structure from motion tasks on images with exposure changes, moving objects, untextured images and more. The paper demonstrates state-of-the-art results on two data sets in both, camera motion estimation and depth estimation. There is no doubt that this paper is a promising step toward solving a structure from motion problem using deep learning . Although it shows some nice results and performance, let us remember that a robust, large scale and accurate deep learning solution is still waiting to be found. This is a hot area of research that will bring more news in the future. Research 8 Research Computer Vision News