Efficient Keypoint Matching in Image Streams <3ms computer vision

Accelerated Computing CUDA CUDA Programming and Performance
1

i just want to present some additional videos showing our realtime image correspondence algorithm.
the GPU time for ~2000 correspondences on a 752x480 image is 2.7ms with a tesla c1060.
note it’s not KLT tracking.

[url=“sidcell longterm correspondences in image streams 01 - YouTube”]sidcell longterm correspondences in image streams 01 - YouTube

[url=“sidcell steps for image correspondence calculation on GPU - YouTube”]sidcell steps for image correspondence calculation on GPU - YouTube

[url=“sidcell ego motion disparity on the GPU - YouTube”]sidcell ego motion disparity on the GPU - YouTube

2

Nice! Do we get to see the source code for this? What algo was mapped onto the GPUs?

3

This is the vehicle they use for testing. The VW Touareg, not the Wolf leading it ;)

http://www.deutschesheer.de/portal/a/heer/…E%2Fcontent.jsp

By the way, this research is done at the university of the German defense forces in Munich.

My take is: No, you won’t get to see the source code. ;)

Christian

4

Do we atleast get to know the algo? Have they published this anywhere? :mellow:

5

of course it’s published:
[Efficient Keypoint Matching for Robot Vision Using GPUs, ICCV09/ECVW]
see computervisioncentral.com

source code is going to get provided, but currently unclear in which form and licence.

greets,
moik

6

direct link for the lazy http://computervisioncentral.com/sites/all…09/AW-9-007.pdf

About the source code: Great news!

7

@cbuchner1: thanks for the link :)

@mikemoik: thanks a lot for the great news about the source code. Eagerly waiting for the moment!! :rolleyes: