ITF Belgium > Program > Speakers  > Luc Van Gool
linkedin twitter facebook email


Luc Van Gool    





Luc Van Gool got a degree in electro-mechanical engineering at the Katholieke Universiteit Leuven in '81. Currently, he is full professor at the Katholieke Universiteit Leuven in Belgium  and the ETH in Zurich, Switzerland. He leads computer vision research at both places, where he also teaches computer vision. He has been a programme committee member of several, major vision conferences, and acted as General Chair of the International Conference on Computer Vision 2011 in Barcelona
and the European Conference on Computer Vision 2014 in Zurich. His main research interests include 3D acquisition and modeling, object recognition, tracking and pose extraction, and the combination of those. Luc Van Gool has received several prizes for his research, including a David Marr Prize in 1998, the ISPRS Helava Award in 2012, and a Koenderink Award in 2016.

He was the coordinator of several European projects, e.g. the ERC Advanced Grant `Variation & the City - VarCity'. His work has led to the foundation of multiple spin-off companies, incl. Eyetronics, GeoAutomation, eSaturnus, EndoSat, kooaba, Procedural, upicto, FaceShift, Fashwell, Parquery, uniqFEED, merantix, Spectando, and Casalva. Several of these have in the meantime been acquired, e.g. by major players like Qualcomm, ESRI, Apple, and Sony.

In 2015 he received the 5-yearly Excellence Prize in Exact Sciences from the Flemish Fund for Scientific Research FWO. He has been and is associate editor of multiple leading computer vision journals. He is editor-in-chief of Foundations and Trends in Computer Graphics and Vision. He lead the Toyota Research labs TRACE-Leuven and TRACE-Zurich, both working on automated cars.


Back to session


Presentation abstract


Looking at evolutions in visual object recognition, one can observe how engineers are reducing their own roles. Initial recognition systems had the engineer decide on all aspects. Features were carefully designed and selected, and a classifier was put together, all manually engineered. But then set in a process of gradual, data-driven automation. First the classifiers were learnt, then the feature selection was automated, and, recently, the very features themselves were automatically generated. So, is there still a role for the engineer? Currently, the DNN architectures are hand-designed. And although one has some overall, intuitive insights – like deeper works better even if one lets the convolutional features shrink proportionally – a theory of DNN design is still beyond the horizon. As DNNs come tantalizingly close to brain structures, it stands to reason to get inspiration from the remaining differences. For instance, biological neural networks make systematic use of feedback, they also seem to solve multiple sub-problems in unison (recognition, motion analysis, depth extraction, …), can learn from weakly annotated data, would not seem to require high precision, and in recognition object hierarchies seem to play a decisive role. We mustn’t parrot biology, but we can definitely learn from it.


Back to program

Back to speakers


Copyright imec 2017