A more or less complete list of my publications is maintained by the university at My publications.  In an attempt to give a more digested version of my recent and ongoing research, I have tried to summarize it in a single heading, which covers most but not all,  namely  “A Statistical Take on 3D Vision“, where statistics should be seen in both meanings , i.e.

Definition of STATISTICS:

1: a branch of mathematics dealing with the collection, analysis, interpretation, and presentation of masses of numerical data.

2:  a collection of quantitative data.


Empirical Evidence


An older version of our 6-axis industrial robots structured light fixture.

As for the second part; I see computer vision as lacking empirical evidence for making good evaluations and experiments. As such I am leading an effort to address this issue, both via our lab, where we are continuously putting a lot of effort into designing equipment and experiments to generate high quality data for computer vision and ground truth to evaluate on. Here a central effort is built around our industrial robot setting. (Please follow the link for relevant publications and freely available data sets).

Acknowledging, that laboratory generated data is needed but not enough, I am also very focused on generating high quality cases based on real cases, mostly from industry. Some of the more recent and prominent found in the MADE project and in the PhD projects of:

Statistical Methods for 3D Computer Vision

Apart from addressing the issue of lacking empirical evidence, I am also interested in generating such data, because I am interested in developing statically based methodology for 3D vision, e.g. 3D model estimation. To do this well, I believe an important factor is having good data available. Currently much of this is ongoing work, where we are e.g. interested in

  • Estimating the radiometric properties of 3d objects from a limited number of images, with Ph.D. student Jannik Boll Nielsen. The issue is, that we are able to estimate geometry of many objects well (given sufficient conditions), but it is very hard for us to make photo realistic renderings of
  • Better regularizing estimated 3D geometry, as a continuation of Vedrana Andersen Dahl’s PhD work. When estimating 3D geometry we often have multiple 3D scans, and noisy data. A way of improving this is via imposing a prior, which removes noise and allows averaging between observations. The standard way of doing this is by minimizint the surface area, which is often not possible. As such we, in Vedrana’s Matser Thesis and PhD work, pioneered using Markov Random Fields on 3D surfaces. Work we (J. Andreas Bærentzen and I) are now revising to get more operational solutions.

Illustration of our ongoing work on regularizing 3D geometry with Markov random fields.

The Cyber Physical Ecosystem for 3D  – ECO3D

To increase the synergy within our section and to promote collaboration between PhD students, such that we have more resources to lift bigger more fundamental research task, my two colleagues, Jeppe Frissvad and Knut Conradsen, and I formed the ECO3D lab/group (please follow the link for details, data  and list of publications).

The general uniting idea of ECO3D is that much of our practical use of information technology, can be illustrated by the following process flow chart

That is, that often want to effect the physical world based on sensory information about it, and that a main use of advanced information technology can be seen in this framework. We thus try to cast projects in this framework, both because it is a great vessel for very useful collaboration, but also because it has implications for the performance metrics we use internally.