Multi-model visual localization

We are describing a visual localization system which is economical, highly accurate, can be applied both indoors and outdoors. In the offline stage, 3D reconstruction and location segmentation is achieved through the use of structure-from-motion techniques on a set of input images. The extracted 3D models constitute our database. In the online stage, user input images are matched against this database and user's camera's location is found. One novelty in this works is the development of new methods to compute the transformation between the 3D model and the target floor plan of the physical neighborhood. These techniques use point-to-point and plane-to line matches. Other novelty is the usage of a multi-part SfM generated model rather than a single one. The approach is justified through experiments. We expect such approach will pave the way for new mobile applications in the field of Augmented reality and Location based services.