Obtaining feature descriptors with SURF

The process of extracting features from an image with OpenCV using SURF is also a single step. It is done by the compute method of our feature extractor. The latter accepts an image and the keypoints of the image as arguments:

key_query, desc_query = self.f_extractor.compute(img_query, key_query)

Here, desc_query is a NumPY ndarray with shape (num_keypoints, descriptor_size). You can see that each descriptor is a vector in an n-dimensional space (n-length array of numbers). Each vector describes the corresponding key point and provides some meaningful information about our complete image.

Hence, we have completed our feature extraction algorithm that had to provide meaningful information about our image in reduced dimensionality. It's up to the creator of the algorithm to decide what kind of information is contained in the descriptor vector, but at the very least the vectors should be such that they are closer to similar keypoints than for keypoints that appear different. 

Our feature extraction algorithm also has a convenient method to combine the processes of feature detection and descriptor creation:

key_query, desc_query = self.f_extractor.detectAndCompute (img_query, None)

It returns both keypoints and descriptors in a single step and accepts a mask of an area of interest, which, in our case, is the complete image.

As we have extracted our features, the next step is to query and train images that contain similar features, which is accomplished by a feature matching algorithm. So, let's learn about feature matching in the next section.