Final Report Summary - HUMANIS (Human Motion Analysis from Image Sequences)
My ERC Starting grant HUMANIS has focused on developing methods for easy and flexible capture of the detailed 3D geometry of time-varying shapes --- such as the face or the full body of an actor, a sports player or even a wild animal --- using a single, low-cost, RGB-only camera, or equivalently existing video footage taken from a single view point. The ambition of the HUMANIS project was to be able to point a single handheld, low-cost commodity camera, at a non-rigid or dynamic scene and be able to capture its detailed 3D geometry from the video stream. Effectively, the main achievement of the HUMANIS project has been to turn a low-cost sensor such as a video camera into a 3D scanner for dynamic scenes. At the start of this project computer vision was very far away from this goal. Dynamic shape capture was only possible by placing markers on the subjects and was confined to studio setups with calibrated multi-camera systems which were costly and technically complex. Some of the most important achievements of the HUMANIS project have been to make possible: 3D reconstruction of non-rigid shapes purely from a video taken with a single camera; 3D reconstruction of generic videos downloaded from youtube depicting complex dynamic scenes; 3D reconstruction of models of object categories from image datasets.