Ask Your Question

surf descriptor accepts only grayscale images?

asked 2013-03-08 07:03:35 -0500

osmander gravatar image

updated 2013-03-08 07:14:36 -0500

Hello all

I have some images with gray values higher than 255 so i keep them in CV_16U type cv::mats. But i realised that surf descriptors need CV_8U images as input. Is this correct?

When i convert them to CV_8U, obviously i am losing some data. So is there a way to get the SURF points for non CV_8U images?


edit retag flag offensive close merge delete

2 answers

Sort by ยป oldest newest most voted

answered 2013-03-08 07:23:09 -0500

updated 2013-03-11 09:31:31 -0500

SURF in OpenCV is the implementation described by the paper: SURF: Speeded Up Robust Feature [Bay et al -]

Comment - Another apporach that could be used is SIFT, which is a sort - like approach to SURF.

Since the paper defines features of grayscale images 8 bit, the implementation does this also. You could dive into the sourcecode and try to create your own SURF implementation for 16bit integers, but it won't be simple.

On the other hand, there are actual reasons for choosing 8bit grayscale images 1) Normalized grayscale images to cope with lightning conditions 2) 8 bit in order to cope with curse of dimensionality

So good luck with trying 16bit, but don't see why this would be usefull. 8 bit already gives you tons of feature points in an image to match. Going to 16 bit will increase this probably BUT will also increase your matching step, making it even more slow then surf already is.

edit flag offensive delete link more


Thanks for the answer. Do you think it is a good idea to map the pixel values of an CV_16U image to 0-255 and then apply SURF detector?

osmander gravatar imageosmander ( 2013-03-08 07:42:06 -0500 )edit

Just a note: SIFT was developed by Lowe et. al not SURF which is by Bay et. al ( @steven: maybe you want to edit your post). Many principals are similar, however they also differ in many steps.

Guanta gravatar imageGuanta ( 2013-03-09 05:31:19 -0500 )edit

Thx guanta for the remark :) Indeed I mixed both of them. Editted post and added both papers as a link.

StevenPuttemans gravatar imageStevenPuttemans ( 2013-03-11 09:32:39 -0500 )edit

answered 2013-03-08 08:01:47 -0500

Yes i think so. Mainly because I have had over more than 100-200 features easily in an image using the surf and sift technique. However, application specific you could go check out other descriptors and feature detectors like ORB, FREAK, FAST, BREAK, ...

edit flag offensive delete link more

Question Tools


Asked: 2013-03-08 07:03:35 -0500

Seen: 2,799 times

Last updated: Mar 11 '13