I am an applied mathematician interested in optimization and modelling, computer vision, deep learning and medical image analysis.
NEWS: We have obtained striking results on the rotation averaging problem! Highly recommended reading: Rotation Averaging and Strong Duality. Joint collaboration with Anders Eriksson at Queensland University of Technology.
NEWS: Our research collaboration with Oxford University (Phil Torr's group) on "Conditional Random Fields Meet Deep Neural Networks" has just resulted in a journal article in IEEE Signal Processing Magazine (special issue on Deep Learning), to appear 2018.
NEWS: The Best Student Paper Award is awarded to "Shape-Aware Multi-Atlas Segmentation" at ICPR 2016, Cancun, Mexico! Read the
NEWS: My PhD Student Petter Strandmark won the Best Nordic PhD Thesis for the years 2013-2014. The prize was announced at SCIA 2015, Copenhagen. Petter, congratulations!
NEWS: The Best Student Paper Award is awarded to our Uberatlas-paper at SCIA 2015, Copenhagen! Read the
Learn more about 3D reconstruction form large-scale image sets and visit Örebro castle at the same time. For more info, read the paper. More reconstructions by Carl Olsson.
We have looked at the problem of extracting one-dimensional structures from large-scale 3D images, for example, automatically segmenting the coronary arties in CT images. For more info, read our paper, Shortest Paths with Higher-Order Regularization, published in IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015.
Research interests include:
Image-based localization. Given a 3D-model of the world, we are interested in determining where an image was taken relative the model.
Multiple View Geometry. While standard algorithms in computer vision suffer from either non-optimality (for example, the 8-point algorithm)
or local minima (for example, bundle adjustment) - or a combination of both - we pursue the goal of globally optimal solutions.
Critical curves and surfaces. In the early 20th century it was noted that for two images of certain scene configurations, there is not a unique solution to the reconstruction problem. We have done a complete classification of all such critical configurations. These results will still be true in 100 years.
Matching and registration. The matching problem is to find corresponding points in different images. One of the main difficulties is to be able to handle large amount of outliers, that is, false matches.
Photometric stereo. The appearance of a surface is
dependent on the incident angle of light. By varying
the light source direction, it is possible to recover
the surface geometry using multiple images from a static camera.
Auto-calibration and critical motions. Auto- (or self-) calibration is the use of calibration constraints, e.g., constant intrinsic parameters, in order to recover Euclidean scene information. However, in certain situations auto-calibration fails to recover unique Euclidean structure, caused by critical motions of the camera.