randomly choose some of your data. !!

for example, randomly choose 50 percent of your data as test_data and the rest will be train_data. 

I think more people can help you if you can clarify the task you are dealing with.

In most cases, people do not directly classify local features. Instead, people encode local features into a single image vector and use this vector as a feature for classification. You may consider first reading papers on local feature encoding, e.g. bag of visual words (BOVW), locally-constrained linear coding (LLC) and Fisher vector.

Then, you have a set of images, each of which is represented by a vector. Suppose we have 10 images. We first assign an ID, from 1 to 10, to each image. Create an array containing 10 integers, again from 1 to 10. Randomly sort the 10 integers using, e.g. matlab function "randperm". Given the sorted list, look at the first 5 integers. Collect the 5 images corresponding to the 5 IDs, and use them as trainig images. Use the remainder as testing images.

If you apply k-Fold Cross Validation, it's a good technique.

 k-fold cross-validation builds on the idea of holdout testing in a clever way by

rotating data through the process. Data is again divided randomly into groups,

but now k equal-sized groups are used.

The train-test process is repeated k times, each time leaving a different segment of

the data out, as the test set.

A common choice for k is 10, resulting in 10-fold cross-validation.

In 10-fold cross-validation, the observations are randomly assigned to 10 groups.

Ten separate models are built and tested on distinct data segments.

The resulting 10 performance measures are unbiased since none of them was built

with test data that was used during training.

The final performance measurement is taken as the mean of these 10

performance measures. The advantage  of this process is that during each fold,

90% of the data is available for training,  

The final performance metric is based on 100%  of the data!

What you really need here is to split the images into a training set and a test set. If you have a resent version of MATLAB with the Computer Vision System Toolbox, you can do that easily using the imageSet object and its parition() method.

According to Andrew Ng's online machine learning course, a rule of thumb would be 60% for training, 20% for validation and 20% for test. Further details is available on coursera, Machine Learning course, section 10, "Advice for applying machine learning".

Use 2-fold cross validation, if you getting trouble then you can use classification learner app that present in Matlab2015b onwards. 

wish you best of luck.