shops and building recognition application

asked 2015-03-17 10:45:34 -0500

Abu Gaseem gravatar image

updated 2015-03-17 12:58:43 -0500

I read about Bag of Visual words strategy used in parallel with SVM classifier , I wonder if this strategy will work with my system shops and building recognition , I think BOW strategy will not work gracefully because the objects are nearly same in structure is'nt it ? . I need some experienced people to guide me to the right way to do such a system before i go in the wrong way . Is there a better strategy than BOW for recognition from that type ? Some of my Datasets . It consist from 3 large building and 23 shops.

obj1

obj2

obj3

obj4

As you can see ,There are a few points to discriminate in my datasets and there's more 20 object .

edit retag flag offensive close merge delete

Comments

It all depends on how your objects look like and if can you extract meaningful local features if BoW is useful or not. If you have structureless objects, then maybe just a contour comparison will do better. Maybe you can add some images of your objects to your post.

Guanta gravatar imageGuanta ( 2015-03-17 12:42:08 -0500 )edit

ok , great Idea

Abu Gaseem gravatar imageAbu Gaseem ( 2015-03-17 12:48:26 -0500 )edit

btw, this one also has a nice kmajority solution, in case you want binary features, like akaze or brisk

berak gravatar imageberak ( 2015-03-17 12:50:21 -0500 )edit

@Abu Gaseem: your problem looks like it should be good doable by means of BoW

@berak: cool, thanks for the link!

Guanta gravatar imageGuanta ( 2015-03-17 15:50:36 -0500 )edit

never feel sorry for asking the why question here !

let me try a 'diplomatic' approach: if you got 25000 images, you'll sincerely gain from the BOW reduction, if you only got 20, - get more data !

(usually with machine learning, you gain a lot with collecting a lot of data, and then cutting it down to the relevant parts, BOW is just another means in this direction)

the flann/brute-force approach is nice, if you only have sparse data / few images to compare

berak gravatar imageberak ( 2015-03-17 16:46:16 -0500 )edit

Yes you are right berak , I was checking the SIFT & SURF with two matcher strategy , Brute Force and FlannBased , I gain good result (inliers) when the two Images are for the same object in my data-sets , in contrast there was an negative inliers in other pairs of Images which are'nt for the same object , the reason for that my data-sets has high texture which is the stone of the buildings as you can see above I posted some training Images also they are too bad , I'm planning to take a new dataset without outliers as possible as i can , I think the BOW will be better on my dataset because you know there's a dictionary of words/features and Features are reduced as you mention , the problem is we cannot predicate how much the produced dictionary are distinctive. I will take with your advics

Abu Gaseem gravatar imageAbu Gaseem ( 2015-03-17 17:59:59 -0500 )edit

also SIFT has high repeatability with my data-sets. Thanks very much @berak ,@Guanta,

Abu Gaseem gravatar imageAbu Gaseem ( 2015-03-17 18:01:43 -0500 )edit

@berak,@Guanta I think you misunderstand the situation .,when you ask me to increase the data-set more than 20 Image. the 23 (3 buildings + 20 shops) is not the number of my data-set/training Images ,but it's the number of classes, each building and shops has a label as you can see the third Image above the system should response with Nour label . I dont think the BOW will be the right choice . I hope my note reach to you guys .

Abu Gaseem gravatar imageAbu Gaseem ( 2015-03-18 18:52:21 -0500 )edit

I think @berak and me understood you quite well, and still: BoW is one way you could solve this problem.

Guanta gravatar imageGuanta ( 2015-03-19 04:14:24 -0500 )edit

ok I will try the BOW . I wonder if you can guide me with how i should capture POI Images ,in context of make zoom in, in order to get rid of the outliers , angle of capturing (i.e variants view for the same POI ) ,and the resolution .my phone can capture photos at min-mum quality 480X640 pixels.

Abu Gaseem gravatar imageAbu Gaseem ( 2015-03-19 13:10:42 -0500 )edit