let I have 128 size vector for 5 keypoints for image I1 and 7 keypoints for image I2.
I have used Sum of square of distance method to calculate distance between keypoints. i.e.
1) i have calculated SSD of first keypoint of I1 with all keypoints of I2.
2) calculated ratio of first best match with second best match.
3) if ratio is > 0.8 then i concluded that keypoint of I1 is not matched with any keypoint of I2. Otherwise I have included number of current keypoint and matched keypoint with its score i.e. its euclidean distance in my final match.
I have tested this for two images of same object.
also tested for two images of different object.
In both cases final result gives some entries of keypoint and their euclidean distance.
I have tested above mention two cases with VLfeat library of matlab. It also generate entries in their 'match' and 'score' while testing two images of same object and different object.
What information of this result will help me to tell that images are matched or not matched?
One more question when we try to match keypoints, Keypoints from background of object also participate in matching process. Is it not going to hamper final match?
As per my understanding as we are trying to match 128 vector which is calculated from 16x16 pixel area around keypoint, There could be possibility of same type of vector can be present in image of different object. then how we can say that object is recognized/matched ?
If your task is image search, you can check the following link:
https://en.wikipedia.org/wiki/Bag-of-words_model_in_computer_vision
The simplest way to compute the similarity between two images is like this. After 3) or vl_ubcmatch(), you can obtain the number of matched keypoints. The similarity can then be defined by the number of matched keypoints divided by the total number of keypoints in the two images. If the object is near-rigid, has a sufficiently large size and has limited viewpoint variation, then in most cases the images of the same object should have a larger similarity than those of different objects.
Directly computing the SSD (Euclidean distance) between all pairs of keypoints may not be a good idea due to its poor efficiency. A faster solution is to use approximated nearest neighbor search like randomized KD-trees etc. I guess VLFEAT provides such functionalities. You may want to check the following link:
http://stackoverflow.com/questions/1500498/how-to-use-sift-algorithm-to-compute-how-similiar-two-images-are
- Keypoints from background of object also participate in matching process. Is it not going to hamper final match?
Yes, it basically hamper the matching result.
- There could be possibility of same type of vector can be present in image of different object. then how we can say that object is recognized/matched?
Yes, this situation happens. The solution of this problem is the same as that of your previous question. You may want to google "spatial verification", "geometric verification", "RANSAC", "Hough voting" etc. You can also check my recent paper on this issue:
https://www.researchgate.net/publication/281445234_Robust_spatial_matching_as_ensemble_of_weak_geometric_relations
Return to my first comment. If your task is image search, BOVW is the first solution you should try. It is much faster than directly computing the SSD (Euclidean distance) between all pairs of keypoints. Maybe you can find some tutorials on BOVW from VLFEAT.
Conference Paper Robust spatial matching as ensemble of weak geometric relations
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing