Simona Kocour

Building 3D maps of scenes from sensor data, especially from images, is a fundamental and long-standing problem in Computer Vision. Most 3D reconstruction algorithms aim to produce 3D models that are as accurate as possible, including fine geometric and texture details. However, there are instances in which a user might not want to share a fully-detailed 3D model. For example, in the context of using 3D maps for cloud-based visual localization, a user might want to remove certain parts before uploading maps to the cloud as to not share private details. The level of detail that a user is willing to share will depend on the situation / application and can vary strongly from just removing a few personal details, over removing certain type of objects, replacing actual texture details with generic textures (e.g., replacing a painting with a generic poster), replacing whole rooms / regions with coarse geometry, to sharing only coarse geometry (e.g., floor plans) for the whole map.

The goal of this thesis is to design and develop a novel type of 3D map representations that gives a user control over the level of detail stored in the map, thus enabling the user to control private information stored in maps. The map representation will be heavily based on machine learning, e.g., will build on neural radiance field-based representations and will heavily rely on 2D and 3D scene understanding. In addition to designing the new representation, showing that