Hey there,

Did you try to use KNN-based (with k=1) classifier? I don't understand "without matching like the case of HOG descriptor"

Thank you Mr Artur for your response, actually i want compare the classification with HOG and the classification with SIFT using the K-NN with K=1, but when i try to use the vl_sift of the vlfeat i have D as vector descriptor with different size for the same size of images, and this is not the case for hog descriptor when i use extractHOGFeatures. that's because the sift is a detector and descriptor so in each image we have a different detected points, and hog is a descriptor.

For this purpose, you need to set a fixed number of key points, however, I don't know if it is possible in SIFT from vlfeat. In addition, you can use "bag of words" or " bag of visual words ".

I will try the concept of bag of words to have one descriptor for each image because even if it is possible to fixe the number of key points we can't have the same key point at each time to compare them.

Thank you sir for your help.

HOG is extracted from cells with defined size, hence the amount of HOG descriptors depends on image and cell size.

1) If you want to have a constant number of SIFT descriptors, you can omit the keypoint detection and extract SIFT descriptors from a grid across the image, which would be comparable to your HOG approach. The method is called dense SIFT, or dSIFT. Check out the corresponding VLFeat implementation [1].

2) Comparing HOG to SIFT, you may also extract HOG descriptors from patches around keypoints detected via the SIFT detector (difference of gaussians).

3) For classification, you require image representations of fixed size. Best way is encoding your descriptors to codebook statistics. There are many methods for this, e.g., using bag-of-visual-words (check out hard- vs. soft-assignment), locality constrained linear coding, VLAD, or Fisher vectors. VLAD and FV encode first and second order statistics, which typically boost the classification accuracy [2].

[1] http://www.vlfeat.org/overview/dsift.html

[2] https://doi.org/10.1371/journal.pone.0170629

Thank you Mr. Marco for you reply it's very helpful, and if you allow me I would like to ask you another question if I want to compare these two methods with another global decriptor which of the three approaches you mentioned will be the most suitable for this case (I think the first since it takes into consideration all the picture).

Thank you sir for your help.

Dear Ilham,

if you want to consider all information within your images, you should extract dSIFT. If your images display lots of clutter (background), you should rather incorporate keypoint detection or some sort of segmentation.

Irrespective of this choice, you'll need image representations of fixed size for training a classifier, e.g., SVM [1]. You can achieve this by computing a set of descriptors from your training images, and compute a codebook by applying a clustering method, e.g., K-means [2] or GMM [3]. Using this codebook, you then compute the image representation by encoding the N descriptors of image I using your codebook. Irrespective of N, this representation is a 1xW vector, where W only depends on your codebook size. Read the VLFeat tutorial pages, they'll help you throughout the process. For additional tweaks and improvements, check out [4].

TL;DR:

1) Use dSIFT for considering the whole image.

2) Extract dSIFT descriptors from your training images and apply clustering --> codebook

3) Per image, encode its descriptors using the codebook --> image representation

4) Train and evaluate a classifier on the image representations.

If this is helpful, consider a recommendation. :-)

Best

Marco

[1] http://www.vlfeat.org/overview/svm.html#tut.svm

[2] http://www.vlfeat.org/overview/kmeans.html

[3] http://www.vlfeat.org/overview/gmm.html

[4] https://doi.org/10.1371/journal.pone.0170629

Thank you sir for your answer as well as for the enriching information.