Doppelganger Buildings

I spotted an interesting paper the other day that tries to disambiguate images of similar structures. This is an interesting task in the domain of 3D reconstruction algorithms because you don't want the wrong images being mapped to the 3D structure.

These buildings, from Figure 1 of the paper, are all different. Can you see why?

So how might you go about building a system that can check if the buildings are the same?

You align, then classify using aligned images and masks.

The authors also collected a new dataset for this task, named Doppelgangers, which contains pairs of buildings with a label. The data and project info can be found here.

Some examples from the dataset.