Computer Vision: integral images and CUDPP

Accelerated Computing CUDA CUDA Programming and Performance
1

the CUDPP api seems interesting for calculation of integral images.

a short introduction on integral images is given here:
[url=“http://www.ais.fraunhofer.de/~surmann/papers/icra2005/node12.html”]http://www.ais.fraunhofer.de/~surmann/pape...005/node12.html[/url]

the idea is, to sum up the intensity values of a digital image N(x,y), so that the new image I(x,y)
holds the sum of intensities in the rectangle N(0,0), N(x,y).

given I(x,y) one could compute the intensity sum of any rectangle in N(x,y) with just 2 adds and 2 subs.

a simple way to implement integral images on a parallel architecture is, first to calc a row-integrated image in parallel, and then column-integrate that image in the same way.

any suggestions doing this with CUDPP?

greetings,
moik

2

Have you looked at the satGL sample that comes with cudpp?

The “sat” comes from “summed area tables,” which is a technique once suggested for use in computer graphics for texture filtering. See “Summed-Area Tables for Texture Mapping,” Crow, F.C. Computer Graphics, 18(3):207-212, 1984.

Jeremy Furtek

3

Hi,

I´m also interested in implemented the integral Image with cuda. Did you manage to do it? Did you use this CUDPP?
My idea is to use the scan algorithm from the examples in CUDA and adapted to what I need.

Thank you.

Laura