Ask Your Question

good sift feature

asked 2013-05-08 03:00:11 -0600

miniME gravatar image miniME
6 2

I am trying to find an objects position with opencv and sift, but the object I try to detect doesn't have any good reliable features. So I thought about sticking something to the object. (Some sort of pattern, like a QR code or so, not an image of another object.)

Is there a good pattern with a lot of sift features which delivers reliable data for object recognition?

delete close flag offensive retag edit

2 Answers

Sort by ยป oldest newest most voted

answered 2013-05-11 04:37:34 -0600

SR gravatar image SR
1241 12 23

updated 2013-05-11 04:38:31 -0600

You mention two different problems:

  1. Reliably find distinctive points on an object.
  2. Describe the object or the neighborhood around these points by discriminative description.

One approach to tackle the first problem (in your context) that is suited for local features is to attach custom markers to the object that yield the maximum response of the corresponding feature detector. See the following publication for a solution: F. Schweiger, B. Zeisl, P. Georgel, G. Schroth, E. Steinbach, N. Navab, Maximum Detector Response Markers for SIFT and SURF, Vision, Modeling and Visualization Workshop (VMV), Braunschweig, November 2009.

The latter problem is more fuzzy but the authors of above publication also provide a way to encode a signature in related manner.

link delete flag offensive edit

answered 2013-05-08 05:42:29 -0600

StevenPuttemans gravatar image StevenPuttemans flag of Belgium
8994 3 29 89

Since the SIFT operator actually searches for regions with high change in gradient information (simply put - a Harris Corner approach) you will need a pattern with much information. Binary patterns are mostly not that interesting to find a lot of SIFT features.

I would suggest to use something with alot of clutter, like this example:

image description

link delete flag offensive edit

Login/Signup to Answer

Question tools


subscribe to rss feed


Asked: 2013-05-08 03:00:11 -0600

Seen: 289 times

Last updated: May 11 '13