Perceiving the world in three dimensions (3D) is essential for an autonomous robot for carrying out tasks such as navigating, interacting with other people or objects, analysing structures and so on. Humans are excellent in estimating relative depth and 3D structure of the contents in a scene through evolution. Achieving similar levels of accuracy through machines forms a fun- damental problem in computer vision and is actively pursued in the community. A large variety of ongoing techniques for 3D reconstruction exist in the literature specific to different applica- tions. For instance, the methods of Structure-From-Motion (SFM) [26, 110] and Simultaneous Localization And Mapping (SLAM) [21, 92] are used for reconstructing an entire environment while dedicated methods exploiting human structure exist for modelling people in 3D from im- ages and videos. In my thesis1, I have focused on mitigating some of the issues faced in SFM and monocular 3D Human modelling.