Computer Vision News - February 2020

2 Summary Research 4 by Amnon Geifman Today, many tech companies work on technologies that aim at providing the user some visual mapping services i.e. AR technologies, HoloLens and many more. These technologies frequently perform localization, based on point clouds. Such point clouds are obtained using Structure from Motion algorithms, after which the images are discarded to preserve privacy. Still, the security/privacy issue is not solved: for example, when a device contains a point cloud of our home, we don’t want it to be possible to retrieve the images that were taken. Today’s paper targets this issue and shows that such technologies can be attacked, proving the very disturbing fact that point clouds retain enough information to reveal scene appearance and compromise privacy. The authors suggest a novel method to retrieve images from the scene using point cloud (and optionally colors and SIFTs). Let’s see how they do all this. Introduction Revealing Scenes by Inverting Structure from Motion Reconstructions Every month, Computer Vision News reviews a research paper from our field. This month we have chosen Revealing Scenes by Inverting Structure from Motion Reconstructions. We are indebted to the authors (Francesco Pittaluga, Sanjeev J. Koppal, Sing Bing Kang and Sudipta N. Sinha), for allowing us to use their images. The full paper is here . "point clouds retain enough information to reveal scene appearance and compromise privacy" Method The input to the model is a point cloud where specifically, one viewpoint is chosen and the 3D points are projected into the image plane. The output of the method is a real synthetic image “taken” from the chosen viewpoint. There are several challenges in this task: first, decide which points in the points cloud are visible and which are