Fill Bounding Boxes in 2D array

Question

I have a 2D numpy array which looks like

array([[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0.]]) `

I want to create bounding box like masks over the 1s shown above. For example it should look like this

array([[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], 
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 1., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 1., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 1., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 1., 0., 0., 0.]])

How can I do it it easily? Also how do I do it if other no.s like 2,3 etc exist but I want to ignore them and the groups are mostly 2.

Answer 1

We have skimage.measure to make life easy when it comes to component labeling. We can use skimage.measure.label to label the different components in the array, and skimage.measure.regionprops to obtain the corresponding slices, which we can use to set the values to 1 in this case:

def fill_bounding_boxes(x):
    l = label(x)
    for s in regionprops(l):
        x[s.slice] = 1
    return x

---

If we try with the proposed example:

from skimage.measure import label, regionprops

a = np.array([[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0.]])

We get:

fill_bounding_boxes(x)

array([[0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 1., 1., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 1., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 1., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 1., 0., 0., 0.],
       [0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 1., 1., 0., 0., 0.]])

Answer 2

While the previous responses are perfectly fine, here's how you could do it with scipy.ndimage:

import numpy as np
from scipy import ndimage

def fill_bboxes(x):
    x_components, _ = ndimage.measurements.label(x, np.ones((3, 3)))
    bboxes = ndimage.measurements.find_objects(x_components)

    for bbox in bboxes:
        x[bbox] = 1

    return x

ndimage.measurements.label does a connected component labelling with the 3x3-"ones" matrix defining the neighbourhood. find_objects then determines the bounding box for each component, which you can then use to set everything within to 1.

Answer 3

There is one solution, but its a little bit hacky and I will not program it for you.

OpenCV - Image processing library, has a algorithm for finding Rectangular contour -> Straight or Rotated. What you may want to do is to transform your array into 2D grayscale image, find contours and write inside the contours your 1s.

Check this image - it is from Opencv DOC - 7.a - https://docs.opencv.org/3.4/dd/d49/tutorial_py_contour_features.html

You would be interested in everything that is inside green lines.

---

To be honest, I think seems to me much easier than programming some algorithm for bounding boxes

Note

Of course you dont really need to do the image stuff, but I think it is enough to use opencv's algorithm for the bounding boxes(countours)

Answer 4

This is an interesting problem. A 2D convolution is a natural approach. However, if the input matrix is sparse (as it appears in your example), this can be costly. For sparse matrix, another approach is to use a clustering algorithm. This extracts only the non-zero pixels from the input box a (the array in your example), and runs a hierarchical clustering. The clustering is based on a special distance matrix (a tuple). Merging happens if boxes are separated by a max of 1 pixel in either direction. You can also apply filter for any numbers you need in the initialization step (say only do for a[row, col]==1 and skip any other numbers, or whatever you wish.

from collections import namedtuple 

Point = namedtuple("Point",["x","y"]) # a pixel on the matrix
Box = namedtuple("Box",["tl","br"]) # a box defined by top-lef/bottom-right

def initialize(a):
    """ create a separate bounding box at each non-zero pixel. """
    boxes = []
    rows, cols = a.shape
    for row in range(rows):
        for col in range(cols):
            if a[row, col] != 0:
                boxes.append(Box(Point(row, col),Point(row, col)))
    return boxes

def dist(box1, box2):
    """ dist between boxes is from top-left to bottom-right, or reverse. """
    x = min(abs(box1.br.x - box2.tl.x), abs(box1.tl.x - box2.br.x))
    y = min(abs(box1.br.y - box2.tl.y), abs(box1.tl.y - box2.br.y))
    return x, y

def merge(boxes, i, j):
    """ pop the boxes at the indices, merge and put back at the end. """
    if i == j:
        return

    if i >= len(boxes) or j >= len(boxes):
        return

    ii = min(i, j)
    jj = max(i, j)
    box_i = boxes[ii]
    box_j = boxes[jj]
    x, y = dist(box_i, box_j)

    if x < 2 or y < 2:
        tl = Point(min(box_i.tl.x, box_j.tl.x),min(box_i.tl.y, box_j.tl.y))
        br = Point(max(box_i.br.x, box_j.br.x),max(box_i.br.y, box_j.br.y))
        del boxes[ii]
        del boxes[jj-1]
        boxes.append(Box(tl, br))


def cluster(a, max_iter=100):
    """ 
        initialize the cluster. then loop through the length and merge 
        boxes. break if `max_iter` reached or no change in length.
    """
    boxes = initialize(a)
    n = len(boxes)
    k = 0

    while k < max_iter:
        for i in range(n):
            for j in range(n):
                merge(boxes, i, j)
        if n == len(boxes):
            break
        n = len(boxes)
        k = k+1

    return boxes

cluster(a)
# output: [Box(tl=Point(x=2, y=2), br=Point(x=5, y=4)),Box(tl=Point(x=11, y=9), br=Point(x=14, y=11))]

# performance 275 µs ± 887 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each)
# compares to 637 µs ± 9.36 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each) for 
#the method based on 2D convolution

This returns a list of boxes defined by the corner points (top-left and bottom-right). Here x is the row number and y is the column numbers. The initialization loops through the entire matrix. But after that we only process a very small subset of points. By changing the dist function, you can customize the box definition (overlapping, non-overlapping etc). Performance can further be optimized (for e.g. breaking if i or j greater the length of boxes within the for loops, than simply returning from the merge function and continue).