My projects

CUDArray - CUDA-based NumPy

CUDArray - CUDA-based NumPy

CUDArray is a CUDA-accelerated subset of the NumPy library. The goal of CUDArray is to combine the easy of development from the NumPy with the computational power of Nvidia GPUs in a lightweight and extensible framework.

Shape index histograms for feature description

Shape index histograms for feature description

with Jacob S. Vestergaard, Anders B. Dahl, Rasmus Larsen

We have developed a new feature descriptor relying on statistics of second order curvature. We have applied this method in two competitions within medical image analysis and have achieved good results.

Mixed reality demo: The pond of illusions

Mixed reality demo: The pond of illusions

with Morten Nobel-Jørgensen, Jannik B. Nielsen, Mikkel D. Olsen, Jeppe R. Frisvad & J. Andreas Bærentzen

A small demo consisting of two computers connected over Ethernet. We create a virtual environment were Microsoft Kinect cameras allow people to communicate and feed fish together!

Visual recognition for product tracking in slaughterhouses

Visual recognition for product tracking in slaughterhouses

with Marchen S. Hviid, Mikkel E. Jørgensen, Rasmus Larsen & Anders L. Dahl

Using off-the-shelf computer vision methods we demonstrate a system for recognizing meat cuts at different points along a slaughterhouse production line. Our results shows that the suggested approach is a promising alternative to the more intrusive methods currently available.

Jet-based local feature description

Jet-based local feature description

with Sune Darkner, Anders L. Dahl & Kim S. Pedersen

We develop a local feature descriptor using higher order image derivatives (aka. image jets). Our main finding is that this type of description is very competitive with traditional local feature description based on gradient orientation histograms (e.g. SIFT).

Unscented Kalman filtering for articulated human tracking

Unscented Kalman filtering for articulated human tracking

with Søren Hauberg & Kim S. Pedersen

We present a system for articulated human tracking and show that unscented Kalman filtering (UKF) allows for less likelihood evaluations compared to particle filtering (PF). Moreover, we achieve a smoother tracking with lower errors using UKF rather than PF.