Object recognition with BOW and SIFT features - general approach

Hello, we are pretty new to these arguments and we need some advice for our project.

The assignement is: to recognize an object in a given dataset. All objects are of the same class and we need only to find the correct instance in the training set. E.g. find type of a box of pasta.

We are currently using SURF features;

At first we find keypoints for every training image .

keypoints = detectSURFFeatures(im, 'MetricThreshold', SURFMetricThreshold)';

and we extract features for every keypoint

trainingImagesPoints(n).features = extractFeatures(im, keypoints);

We use a new photo for the test phase:

for that image we find keypoints and the corrisponding features and we search for corrispondences with

pairs = matchFeatures(testFeatures, trainingImagesPoints(i).features, 'MatchThreshold', 10);

In the end we count how many pairs we got and the training image's label with more matches is the candidate one to be the answer of the recognition.

This is a very simple solution; the question is:

we are currently using only one training image for every label, they're not many. Can a Bag of words approach be useful our problem? We know it is used for classification purposes, what about recognition? We can take multiple pictures of the same object, find keypoints, features, create clusters and histograms and in the end use a classifier to get an answer.