Revision history [back]

dear FLY, do yourself a favour, leave that code alone for a while.

imho, you got 2 major problems there:

• seems you can't code your way out of a paper bag. (sorry for the harsh truth, but unless you solve that, this won't go anywhere)
• you're totally impatient. you're still having problems processing your data for a svm, yet you're messing with bow.

i can only try to address the 2nd problem

you don't have to resize the images. surf is scale invariant

you've got to find a way how to properly use an SVM with your descriptors:

extractor->compute( row_img, keypoints, descriptors_1);

descriptors_1 will be a Mat with n rows, one for each descriptor found in the image. n will vary for each image(it'll find 3 descriptors in one image, 5 in another), so you'll have to find a way dealing with that, either :

• some procrustes approach( limit it to a fixed number of descriptors, and ignore images that give less ), then reshape(1,1), that desc-mat to a single row [there's your runtime error, btw]
• go for single descriptors in the first place, chop the descriptor-mat into rows, and insert single descriptors ( and a label for each ) this would avoid the problem of different desc. counts, but add a problem in the prediction, you'll have to do predict on several single descs as well, and find some kind of voting scheme.

whatever you choose, you'll need the same procedure in the prediction stage, btw.

so , again. start all simple this time, take only 2 positive and 2 negative images, and try to train an SVM only on that. (maybe you even should skip the images/descriptors, and start with simple mockup data, [2,5,3,4] : label 1; [3,5,3,1] : label -1 , etc )

don't proceed, until you get that step right, please.

dear FLY, do yourself a favour, leave that code alone for a while.

imho, you got 2 major problems there:

• seems you can't code your way out of a paper bag. (sorry for the harsh truth, but unless you solve that, this won't go anywhere)
• you're totally impatient. you're still having problems processing your data for a svm, yet you're messing with bow.

i can only try to address the 2nd problem

you don't have to resize the images. surf is scale invariant

you've got to find a way how to properly use an SVM with your descriptors:

extractor->compute( row_img, keypoints, descriptors_1);

descriptors_1 will be a Mat with n rows, one for each descriptor found in the image. n will vary for each image(it'll find 3 descriptors in one image, 5 in another), so you'll have to find a way dealing with that, either :

• some procrustes approach( limit it to a fixed number of descriptors, and ignore images that give less ), then reshape(1,1), that desc-mat to a single row [there's your runtime error, btw]
• go for single descriptors in the first place, chop the descriptor-mat into rows, and insert single descriptors ( and a label for each ) this would avoid the problem of different desc. counts, but add a problem in the prediction, you'll have to do predict on several single descs as well, and find some kind of voting scheme.

whatever you choose, you'll need the same procedure in the prediction stage, btw.

so , again. start all simple this time, take only 2 positive and 2 negative images, and try to train an SVM only on that. (maybe you even should skip the images/descriptors, and start with simple mockup data, [2,5,3,4] : label 1; [3,5,3,1] : label -1 , etc )

don't dare to proceed, until you get that step right, please.

dear FLY, do yourself a favour, leave that code alone for a while.

imho, you got 2 major problems there:

• seems you can't code your way out of a paper bag. (sorry for the harsh truth, but unless you solve that, this won't go anywhere)
• you're totally impatient. you're still having problems processing your data for a svm, yet you're messing with bow.

i can only try to address the 2nd problem

you don't have to resize the images. surf is scale invariant

you've got to find a way how to properly use an SVM with your descriptors:

extractor->compute( row_img, keypoints, descriptors_1);

descriptors_1 will be a Mat with n rows, one for each descriptor found in the image. n will vary for each image(it'll find 3 descriptors in one image, 5 in another), so another).

since SVM expects all training vec to have the same length, you'll have to find a way dealing with that, either :

• some procrustes approach( limit it to a fixed number of descriptors, and ignore images that give less ), then reshape(1,1), that desc-mat to a single row [there's your runtime error, btw]
• go for single descriptors in the first place, chop the descriptor-mat into rows, and insert single descriptors ( and a label for each ) this would avoid the problem of different desc. counts, but add a problem in the prediction, you'll have to do predict on several single descs as well, and find some kind of voting scheme.

whatever you choose, you'll need the same procedure in the prediction stage, btw.

so , again. start all simple this time, take only 2 positive and 2 negative images, and try to train an SVM only on that. (maybe you even should skip the images/descriptors, and start with simple mockup data, [2,5,3,4] : label 1; [3,5,3,1] : label -1 , etc )

don't dare to proceed, until you get that right,

please.

dear FLY, do yourself a favour, leave that code alone for a while.

imho, you got 2 major problems there:

• seems you can't code your way out of a paper bag. (sorry for the harsh truth, but unless you solve that, this won't go anywhere)
• you're totally impatient. you're still having problems processing your data for a svm, yet you're messing with bow.

i can only try to address the 2nd problem

you don't have to resize the images. surf is SURF is scale invariantinvariant, just trust in it.

you've got to find a way how to properly use an SVM with your descriptors:

extractor->compute( row_img, keypoints, descriptors_1);

descriptors_1 will be a Mat with n rows, one for each descriptor found in the image. n will vary for each image(it'll find 3 descriptors in one image, 5 in another).

since SVM expects all training vec vecs to have the same length, you'll have to find a way dealing with that, either :

• some procrustes approach( limit it to a fixed number of descriptors, and ignore images that give less ), then reshape(1,1), that desc-mat to a single row [there's your runtime error, btw]
• go for single descriptors in the first place, chop the descriptor-mat into rows, and insert single descriptors ( and a label for each ) this would avoid the problem of different desc. counts, but add a problem in the prediction, you'll have to do predict on several single descs as well, and find some kind of voting scheme.

whatever you choose, you'll need the same procedure in the training and prediction stage, btw.stage.

so , again. start all simple this time, take only 2 positive and 2 negative images, and try to train an SVM only on that. (maybe you even should skip the images/descriptors, and start with simple mockup data, [2,5,3,4] : label 1; [3,5,3,1] : label -1 , etc )

don't dare to proceed, until you get that right,

please.