Deep representations of structures in the 3D-world

This thesis demonstrates a collection of neural network tools that leverage the structures and symmetries of the 3D-world. We have explored various aspects of a vision system ranging from relative pose estimation to 3D-part decomposition from 2D images. For any vision system, it is crucially important to understand and to resolve visual ambiguities in 3D arising from imaging methods. This thesis has shown that leveraging prior knowledge about the structures and the symmetries of the 3D-world in neural network architectures brings about better representations for ambiguous situations. It helps solve problems which are inherently ill-posed.