Damascene

Damascene is an implementation of the gPb image contour detector. By reevaluating components of the gPb algorithm in a parallel context, and creating an eigensolver specially suited to Normalized Cuts image segmentation, we have made high-quality image contour detection more efficient. Damascene is a parallel implementation of these ideas in CUDA. On an Nvidia GTX280, we can perform image contour detection 114x faster than the original serial CPU gPb detector (which doesn't use our algorithmic tweaks), while still providing the same accuracy on the contour benchmark of the Berkeley Segmentation Dataset. For more details, please reference our upcoming ICCV 2009 paper.

Download Damascene here: Damascene 0.1

Acknowledgements

Bor-Yiing Su
Narayanan Sundaram
Yunsup Lee
Mark Murphy
Kurt Keutzer
Michael Maire
Pablo Arbelaez
Mark Hoemmen
Jitendra Malik
Jim Demmel


The ParLab @ Berkeley The Gigascale Systems Research Center	Damascene Damascene is an implementation of the gPb image contour detector. By reevaluating components of the gPb algorithm in a parallel context, and creating an eigensolver specially suited to Normalized Cuts image segmentation, we have made high-quality image contour detection more efficient. Damascene is a parallel implementation of these ideas in CUDA. On an Nvidia GTX280, we can perform image contour detection 114x faster than the original serial CPU gPb detector (which doesn't use our algorithmic tweaks), while still providing the same accuracy on the contour benchmark of the Berkeley Segmentation Dataset. For more details, please reference our upcoming ICCV 2009 paper. Download Damascene here: Damascene 0.1 Acknowledgements Bor-Yiing Su Narayanan Sundaram Yunsup Lee Mark Murphy Kurt Keutzer Michael Maire Pablo Arbelaez Mark Hoemmen Jitendra Malik Jim Demmel