Our latest research activities involve around topics in 3D computer vision and machine learning.
Modern deep learning produces highly accurate results on many challenging datasets. State-of-the-art methods are, however, often not directly transferable to real-time applications or embedded devices. We build deep convolutional neural networks especially tuned for tasks that require real-time processing and memory efficiency.
The problem of online structure from motion is also known as simultaneous localisation and mapping. It involves a system that estimates a sensor’s pose and the structure of the environment in real-time. We develop methods for visual odometry, loop closure detection and pose graph optimization.
Energy-minimization methods are ubiquitous in computer vision and related fields. In our work we design efficient low-level optimization strategies at pixel level, and robust optimization methods for 3D vision problems with high complexity.
Joint object recognition and pose estimation is an important task in robotics applications and in automated manufacturing environments. In past work we established fast algorithms for the detection of rigid objects in cluttered point cloud data, and non-rigid human body shape estimation from range camera depth.
In collaboration with CRL’s Speech Technology Group we produced a complete system for expressive visual text-to-speech. Our system is able to producing expressive verbal and visual output in the form of a 'talking head', given an input text and a set of continuous expression weights.
Automated change detection is very valuable for inspection and maintenance. We conducted fundamental vision research to facilitate low cost change detection systems. Specifically, by combining 3D modelling with image registration, we compare images from different time instances and visually enhance change.
In 2007, CVG started to extend 3D reconstruction techniques and applied them to faces and the human body in motion. Since then, CVG has created several computer vision technique using multiple cameras and known light sources to capture static and deforming surfaces at high levels of detail.
Building 3D models of static scenes needs to be very accurate to be usable, for example, to create 3D prints of object. We developed multi-view geometry techniques to reconstruct a 3D shape at high precision.