Hello,
I would like to develop, for academic purpouse, an object recognition system that uses (among other things) edges or corners matching.
Basically I extract corner information on my model image then i try to find this object inside a query image using corner information.
Do you have any hint?
You may be interested in a corner descriptor called PAS (Pair of Adjacent Segments). It can be computed using Canny edges and it is invariant to translation and scale changes. See the paper for details and how to use it for object detection:
For edge-based object recognition and 6-DOF pose estimation see the papers:
If you don't need 6-DOF pose and you need just recognize and localize an object in an image then take a look at the Line-2D algorithm:
Its implementation is available in OpenCV in the objdetect module. It uses orientations of gradients on a boundary of an object to create Line-2D templates so it is essentially recognition by silhouettes.
No, it doesn't has online documentation yet but it has inline doxygen docs and a c++ sample. The interface is located in objdetect.hpp in cv::linemod:: namespace.
Ilya, did I got it right, that you would suggest Line-2D for doing things like the Shot Note iOS app does (http://dl.dropbox.com/u/2331404/shootnote.jpg). Basically it uses shapes on the corners to recognize the position of the paper. I want to do just the same thing, but kinda stuck with the methods to use. Tried to use SURF, but it doesn't take into consideration the shape of the corner objects and doesn't seem to work great at my sample images.
Yes, Line2D should work but you need to train it on rectangles of different proportions. Also you need to choose the largest detected quadrangle as the final answer or at least to normalize detection score by size of a quadrangle to exclude detections inside of the note.
Hi, I just find this video Real-time Learning and Detection of 3D Texture-less Objects and also the paper, it is quite recent.
Asked: 2012-07-27 15:45:03 -0600
Seen: 8,189 times
Last updated: May 16 '13
Hi,
can you be more specific and tell us why do you want to use edges/corners instead of the most common approach which is to use interest points (e.g. SIFT)?
Because, unfortunately my prof., obbligated me to do so. I have already implemented a system using ORB and it works fine, now i need to do this using corner or edge. A reason could be that SIFT/SURF doens't take in account the shape of the object
The question is good. Sometimes you don't have features, i.e. objects without texture, or transparent objects. There are some approaches that analyze contours, but I'm not specialist, let's wait for gurus :) In the meantime you can start from this paper: http://www.roboticsproceedings.org/rss08/p35.html, it is about one particular problem, but may be it'll inspire you!
thanks for the paper. Anyway let's hope for some Guru response :D