Hi guys,
I have an image of a box and I need an algorithm that does nothing more than to find the box and crop to its dimensions. If possible regardless of the background but as this will work poorly the algorithm should assume a fairly monotone background.
I have been using histogram analysis which isn't sufficient. I thought of adding edge detection but it doesn't seem too helpful. Since rectangle detection is quite the simple problem definition, I was wondering what the latest research on this is, anyone have any insights, I've been away from the area for a few years.
Thanks.
-CProgrammer
image segmentation
Here's a resource that might help. You can get a fairly complete draft of this book: http://szeliski.org/Book/
This is good, thanks! Had been considering level sets, and this seems to confirm that choice.
Hm may opt to implement this myself, but anyone know of an existing solution, am willing to pay for it, for linux. I want this done quickly. Also anyone have experience using OpenCV and can recommend it for this?
-CProgrammer
Hm may opt to implement this myself, but anyone know of an existing solution, am willing to pay for it, for linux. I want this done quickly. Also anyone have experience using OpenCV and can recommend it for this?
-CProgrammer
Hough transform?
Find edges, filter the candidate set by rectangular sets of lines. If there are too many candidates, minimize the error to shapes that form rectangles or cubes. If available, using reverse camera transform may allow fitting in 3D model space.
Find edges, filter the candidate set by rectangular sets of lines. If there are too many candidates, minimize the error to shapes that form rectangles or cubes. If available, using reverse camera transform may allow fitting in 3D model space.
Hey Antheus,
thanks for the reply.
I've thought of Hough transform, I'm worried though it may not suffice.
EDIT: The issue will likely lie in the fact that the rectangles can be slightly oval shaped. This means the algorithm will likely still need to segment further. The final touches could be done using histogram analysis or another edge detection algorithm but that would defeat the purpose of the Hough transform.
I'll give it a try. Come to think of it. What I should really have is an application that can load some of my sample images and then apply all kinds of algorithms on it. From graph based, level sets, hough etc.
Anyone know of one.
-CProgrammer
thanks for the reply.
I've thought of Hough transform, I'm worried though it may not suffice.
EDIT: The issue will likely lie in the fact that the rectangles can be slightly oval shaped. This means the algorithm will likely still need to segment further. The final touches could be done using histogram analysis or another edge detection algorithm but that would defeat the purpose of the Hough transform.
I'll give it a try. Come to think of it. What I should really have is an application that can load some of my sample images and then apply all kinds of algorithms on it. From graph based, level sets, hough etc.
Anyone know of one.
-CProgrammer
Nice book, thanks for the link!
I would recommend segmentation over edge detection because it can make use of more information.
Mean shift is good (search terms: edison comaniciu), graph-based segmentation (graph-based segmentation felzenszwalb) is a good bit faster.
Both of those have demonstrators or code you can try out. Depending on the radiometry, you might also get away with a gradient filter+MSER.
HTH+HAND
Jan // on the road
I would recommend segmentation over edge detection because it can make use of more information.
Mean shift is good (search terms: edison comaniciu), graph-based segmentation (graph-based segmentation felzenszwalb) is a good bit faster.
Both of those have demonstrators or code you can try out. Depending on the radiometry, you might also get away with a gradient filter+MSER.
HTH+HAND
Jan // on the road
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement