Aligning and Updating Building Maps with Aerial Images by Multi-Task, Multi-Resolution Deep Learning
A large part of the world is already covered by maps of buildings. However when a new image of an already covered area is captured, it does not align perfectly with the existing map, due to a change of capture conditions, and errors in the map data. Those deformations can only be partially corrected, which leads to misalignments. Leveraging multi-task learning, our model aligns the existing buildings to the new image through a displacement output, and detects new buildings that do not appear in the map through a segmentation output. We also apply our method to buildings height estimation, by aligning to a pair of stereo images. An extension of our work in the noisy-supervision setting allows for the alignement of misaligned building maps without the use of perfect ground truth data for training.
Jean-Philippe BAUCHET (TITANE)
Title Kinetic shape approximation
Kinetic data structures consist in a set of geometric primitives whose coordinates are continuous functions of the time. The purpose of such frameworks is to maintain the validity of a set of statements that apply to such primitives over time. In this talk, I will present how kinetic data structures can be used in a context of image partitioning and surface approximation from a point cloud.