Jhacson Meza

Building 3D maps of scenes from sensor data, especially from images, is a fundamental and long-standing problem in Computer Vision. 3D maps play an important role in many applications. They are used to localize and guide users in Augmented Mixed / Virtual Reality applications, they are used for path planning by robots, and they are used to document cultural heritage. 3D mapping algorithms are also used for content creation in the entertainment industry, e.g., to digitize objects and scenes for games and movies. Furthermore, there is a lively community sharing 3D models constructed from images on the Internet.

State-of-the-art mapping algorithms mostly assume that the scene is static and does not change during the data capture process. Recent works based on neural implicit scene representations showed that it is possible to model illumination changes, e.g., day-night changes, if they are contained in the captured data. Yet, the appearance of real-world scenes not only varies due to illumination changes, but also due to geometric changes. For example, vegetation grows over seasons in outdoor scenes, and objects move around in indoor scenes. This type of scene dynamics cannot be modeled by current 3D mapping algorithms.