ECCV 2020 Daily - Thursday

2 Oral Presentation 4 Marcel Geppert is a PhD student at the Computer Vision and Geometry group at ETH Zurich, under the supervision of Marc Pollefeys. His work addresses the significant privacy concerns that arise as localization and mapping solutions becomemore cloudbased. He speaks to us ahead of his oral presentation today. Current structure-from-motion (SfM) methods use a lot of image data, and with new methods for mobile devices processing images on a server in the cloud, private user information could be revealed . This work aims to remove as much of this information as possible, while maintaining the same results in the form of a point cloud or map. “Usually, we would not transfer images directly, we would do feature extraction like SIFT,” Marcel explains. “We would use a deep learning method by Francesco Pittaluga that takes those features and their keypoint positions and gets surprisingly good results by just reconstructing the original image. In our method, instead of keeping the exact 2D keypoint position, we add a degree of freedom and write it as a line that includes the original extracted keypoint. The position information is missing, so it is not possible to reconstruct the input image anymore.” The inspiration for this work came from co-author Pablo Speciale . He was first author on twopapers: PrivacyPreserving Image Queries for Camera Localization and Privacy Preserving Image-Based Localization . The first proposes toprotect the map with respect to the user. Instead of having a point cloud, it does the same thing in 3D, but replaces each point with a line . With this method, it is still possible to estimate the camera position based on point features in the image and lines on the map. The second paper has a point cloud with line features – which is also used in this work – and localization of that. The problem with this method is that to create the map, you still need to use standard SfM methods . “That is basically what we added here, ” Marcel tells us. “The single modules that we use were known before. The localization was done by Pablo. Point triangulation is not that hard. You need more views than for standard keypoints, but with lines it is still possible. The Privacy Preserving Structure-from-Motion DAILY T h u r s d a y