My research in the area of 3D Computer Vision has consistently focused on the inference of 3D information from video. The core underlying question driving my research can be formulated as: to what extent is it possible to obtain detailed 3D models just by analysing the raw input given by a standard video camera?
While my research first focused on recovering the 3D structure of static scenes containing rigid objects, my attention has turned to the problem of inferring the detailed 3D shape of non-rigid objects from the footage acquired with a single standard video camera.

In November 2008 I was awarded an ERC Starting Independent Researcher Grant (HUMANIS) to focus on the problem of reconstructing 3D models which represent the full geometry of deforming and articulated objects, such as the human body, but in particular in acquiring them automatically and only from the stream of images acquired with a single conventional camera, rather than using multiple-camera setups, specialised sensors (such as depth cameras), prior knowledge about the objects to be reconstructed or training data – a purely data-driven approach.

The research carried out by my team provides robust solutions to the most challenging aspects of non-rigid and articulated 3D inference from images: modeling highly deformable surfaces, such as flexible cloth; reconstructing full human body motion simply from 2D matching data; 2D deformable tracking; dense optical flow estimation and non-rigid video registration. Recent breakthroughs include: (i) the first dense approaches to the 3D reconstruction of non-rigid or dynamic scenes where every pixel in the image is reconstructed in 3D and (ii) a novel unified approach that simultaneously segments the scene into different object parts while reconstructing them in 3D. This allows the acquisition of semantic models for instance of a human body where each point is assigned to a body part corresponding to the underlying skeletal structure such as forearm, torso, or head.