NumPy

What is NumPy

NumPy is a scientific computation package

It offers many functions and utilities to work with N-Dimension arrays

Largely used by other libraries such as OpenCV, TensorFlow and PyTorch to

deal with multi dimensional arrays (e.g., tensors or images)
How to install

We can easily install NumPy using pip by running

pip install numpy

Then, in our python script we have to import it

N-Dimensional array

N-Dimensional array are, for instance, the followings:

1-D 2-D 3-D 4-D

NumPy array: how to create an array

In order to create an array, we can use the array function, passing a list of
values and optionally the type of data

NOTE: NumPy arrays must be homogeneous, so each element must have the same type

NOTE: notice that if the type is not set, NumPy will decide the type for you. Default value for NumPy
arrays is Float64
NumPy array: how to create an array

Moreover, NumPy offers standard functions to easily create arrays. Some of

them are ones, zeros, ones_like, zeros_like and eye, but there are many more

We use the functions zeros and ones to create an array with given shapes in
which each element is, respectively, 0 or 1
NumPy array: how to create an array

Given a NumPy array with a certain shape, zeros_like and ones_like allow to
create a 0 and 1 arrays with the same shape
NumPy array: how to create an array

Using the function arange(start, stop, step), we obtain a NumPy array

containing all elements from start to stop, using a step spacing consecutive
elements

Notice that the values are generated within the half-open [start, stop[, so stop is
not included
NumPy array: how to create an array

With eye we create an identity matrix, so a matrix full of zeros except for the
diagonal
[[1. 0. 0. 0. 0.]
[0. 1. 0. 0. 0.]
[0. 0. 1. 0. 0.]
[0. 0. 0. 1. 0.]
[0. 0. 0. 0. 1.]]

Sometimes we need random values. We can obtain an array filled with random
values calling the rand function
NumPy array: how to create an array

Finally, we can create an array that contains a single scalar value using full

We are able to obtain the same result using ones/ones_like function

NumPy attributes

Each array has got attributes, such as dtype or shape. Attributes contain
important information related to that particular array

dtype allows to know the type of the array

NumPy attributes

shape give you back the size of the array along each dimension
Changing the shape of arrays

Given an array, we can add a new dimension using expand_dims

Changing the shape of arrays

A common operation consist in changing the shape of a given array. For

instance, we can turn a 10 elements array into a 2x5 using the reshape function

We can “ask” NumPy to complete by himself the shape, using -1

Changing the shape of arrays

Notice that this operation is valid for just 1 dimension. In fact, If more
dimensions are unknown, NumPy will throw ValueError
Changing the shape of arrays

Using squeeze, we can remove all the single dimensional entries of the array

However, squeeze allows also to specify the axis to delete (scalar, tuple or
None. Default is None)
Elements of arrays

Given an array, you can access to its elements by index notation

Elements can be retrieved also using item function

Elements of arrays

Slice notation (the same used for python strings) is valid also for arrays
Elements of arrays

Given two array, we can concatenate them together to obtain a single array as
output thanks to the concatenate function
NumPy math

Since NumPy is a scientific package that offers easy and even complex
functions that you can apply to arrays.

In the following, we are going to see some of them

NumPy math: sum and subtraction

Given two arrays, you can sum or subtract them just using + and - operators

In this case, both arrays have the same shape, so the operations are performed
element-wise
NumPy math: broadcasting

Sometimes, our arrays have not the same shape, but NumPy is smart enough
and try to “fit” the arrays. This operation is called broadcasting

Notice that y is a scalar, but both array_sum and array_sub have shapes (4,2,3)
NumPy math: broadcasting

Broadcasting can’t work for all the cases: when operating on two arrays,
NumPy looks at their shapes. The shapes are compatible if, in the element-wise
comparison, they are equals or one dimension is 1. The resulting shape is the
maximum shape along each dimension.

If broadcasting can’t be applied, ValueError would be raised

NumPy math: multiplication

Given two NumPy arrays, we can perform element-wise multiplication using *

or multiply function
NumPy math: matrix multiplication

Given two NumPy arrays, we can perform matrix multiplication using matmul
function

For N-Dimensional arrays (N>2), matmul applies broadcasting, treating the

array as a stack of matrices
Conditions

We can use where function to apply a condition.Given an input array, a

condition and two arrays x and y, for each element we sample from x if the
condition is verified, from y otherwise
Conditions

Notice that NumPy is quite optimized, so when possible try to used “native”
NumPy way instead of other approaches
NumPy Input/Output

NumPy provides a set of functions to write and read directly from the
filesystem

Plain text (.txt) and csv (.csv) can be loaded using loadtxt function, providing
the path to the file, the data type and the delimiter
NumPy Input/Output

We can store NumPy arrays in two ways:

● binary file: using np.save we are able to serialize our arrays in the local file
system. We will obtain a .npy file containing the array
● txt: using savetxt function we will store our 1D or 2D array in a new txt

Finally, .npy files can be read using np.load function

Plots

Sometimes, we want to display values in a human readable form. Suppose for

instance you have got temperature values collected by a sensor, one
measurement per hour for a week. Display such values in a chart would
simplify largely the reading!

We can visualize NumPy arrays using some chart libraries, like Matplotlib
Matplotlib

Matplotlib is an open source plotting library able to produce high quality graphs
and charts. Easy to install using pip (just “pip install matplotlib”)

It offers a large set of plot types (e.g., histogram, scatter, line, 3D and more),
and uses NumPy arrays to handle data

It can run also on interactive environments such as Jupyter

Matplotlib

Given two collection of values, m1 and m2, we can visualize them using
Matplotlib
OpenCV
OpenCV

OpenCV is an open source Computer Computer Vision library. It allows to

develop complex Computer Vision and Machine Learning applications fast,
offering a wide set of functions.

Originally developed in C/C++, now OpenCV has handlers also for Java and
Python

it can be exploited also in iOS and Android apps.

In Python, OpenCV and NumPy are strictly related

OpenCV

In particular, some of the functions offered by OpenCV are:

● Image Handling (read an image, write an image etc)

● Corner Detection (Harris, Shi-Tomasi etc)
● Camera Calibration
● Features Detection and Description (ORB, SIFT, SURF etc)
● K-Nearest Neighbour
● Depth estimation (Block Matching, SGM etc)
● Optical Flow (Lucas-Kanade)

and many more

OpenCV: installation

We can install OpenCV directly by pip, calling

pip install opencv-python

Then, in our Python script, we can import it as follows:

OpenCV: Image Handling

As we said, OpenCV offers functions to read and write images.

We can open an image using the imread function:

Moreover, it is able to handle various image format (png, jpeg etc) and data
types (8bit, 16 bit etc)
OpenCV: Image Handling

Once opened, OpenCV returns a numpy array that stores the image (each value
of the array is a pixel)

OpenCV default format is BGR, so we have to swap the first and the last
channels in order to manage a RGB image. We can do it manually or invoking
the cvtColor function

cvtColor helps in converting colored images (BGR or RGB) to grayscale just

using as options cv2.BGR2GRAY or cv2.RGB2GRAY
OpenCV: Image Handling

(A) (B)

In figure (A), the original RGB image, while in figure (B) the same picture saved
using BGR format
OpenCV: Image Handling

(A) (B)

We can use cv.cvtColor function also to obtain grayscale images

OpenCV: Image Handling

Instead, we can write an image in our file system using the imwrite function

However, remember that OpenCV expects a BGR image, so if img is a RGB you
must convert to BGR using cv2.cvtColor
OpenCV: Image Handling

OpenCV allows to resize images using the resize function. It takes the image
and the new shape

(A) Original image

(B) Image resized at 320x240

OpenCV: filters

2D Convolution in OpenCV is straightforward: you have just to call the filter

function, passing as input the image and the kernel
OpenCV: filters

Changing the filter, we would obtain different results. For instance, high-pass
filter can be obtained through a zero-sum kernel

-1 0 1

-2 0 2

-1 0 1
OpenCV: filters

We can obtain the same result* using separable filters

2
* -1 0 1
* 1
Example: Stereo Matching
from Middlebury Dataset

FAR CLOSE
 Example: 3D reconstruction
point cloud can be visualized using MeshLab

from Middlebury Dataset

Exercises
Exercise 1

Given the image canyon.png load it using OpenCV, split the channels and save
each channel in a new image called as the channel.

For instance, the red channel have to be saved as red.png

Exercise 2

Using the same image of the previous exercise, load it as gray-scale and
replace all pixels with intensity lower than 80 with 0, 1 otherwise. Save it both
as a new image, called mask.png, and as npy. Finally, apply the mask to the
original image, keeping the original value where the mask is 1, 0 otherwise
Exercise 3

Using Matplotlib, display intensity values of canyon.png (loaded as grayscale

image) in a bar chart. In particular, for each intensity value, the height of the
column is the number of pixels that have such intensity value