Ask Your Question
1

Image matching in a large image collection

asked 2014-05-17 05:10:44 -0600

still_learning gravatar image

updated 2019-12-09 08:04:16 -0600

Akhil Patel gravatar image

For some time now, I've been trying to make a detection system for detecting Magic the Gathering cards.

I've tried the most naive method. For each image -> keypoints -> descriptors -> matcher. On completion, match a collection of pictures taken from a smartphone. (I'm not trying to do the matching on a phone. I send the image to the PC, then return the results back)

I've tried every single descriptor/matcher setting combination. The results are good but not great. I get a lot of false positives. What else can I do to achieve better results?

edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted
2

answered 2014-05-17 17:31:17 -0600

How many cards are in your collection? Have you masked you cards so that you only match the image in the upper part of the card and not the whole card? The border is always similar so that it can't be used to discriminate.

Do you have any sample images? Or visualizations of false positives.

Have you checked the Bag of Word implementations? One important paper is http://www.vis.uky.edu/~stewe/publications/nister_stewenius_cvpr2006.pdf

edit flag offensive delete link more

Comments

There is about 26.000 images. I use only cropped art for detection. The best results were achieved with FAST/FREAK combination or SURF/SURF. I converted the descriptors to float and used flann with kmeans params when working with binary descriptors.

I have sample images, but no visualization of false positives.

I checked BoW, but I work mainly with binary descriptors due to the fact SURF/SIFT being non free. I had no success implementing this method yet due to errors i'm getting.

still_learning gravatar imagestill_learning ( 2014-05-18 07:58:25 -0600 )edit
1

You need to group your local descriptors to a global one aka use some form of BoW to get good results. This is also possible with binary features, see http://answers.opencv.org/question/24835/is-there-a-way-of-using-orb-with-bow/#24858 .

An alternative way would be to use global descriptors in the first way, like the GIST descriptor, see http://people.csail.mit.edu/torralba/code/spatialenvelope/ .

Guanta gravatar imageGuanta ( 2014-05-18 08:52:06 -0600 )edit

Thank you for your help.

still_learning gravatar imagestill_learning ( 2014-05-18 10:02:10 -0600 )edit

i think u can use template matching it is implemented on opencv http://docs.opencv.org/doc/tutorials/imgproc/histograms/template_matching/template_matching.html

AZAZA Jihed gravatar imageAZAZA Jihed ( 2014-05-19 08:46:16 -0600 )edit

Question Tools

Stats

Asked: 2014-05-17 05:10:44 -0600

Seen: 662 times

Last updated: May 17 '14