OpenCV moments() returns zero, causing ZeroDivisionError in calculation of object center and area

Question

I am trying to compute the area and the center coordinates of objects in a binary mask using opencv. However, I noticed that in some cases I get the wrong result. For example, if I get the contours like this:

import numpy as np
import cv2

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

contours, _ = cv2.findContours(
    binary_mask.astype(np.uint8),
    cv2.RETR_EXTERNAL,
    cv2.CHAIN_APPROX_SIMPLE)

>>> contours
(array([[[0, 3]],

        [[5, 3]]], dtype=int32),
 array([[[2, 0]],

        [[2, 1]],

        [[3, 1]],

        [[3, 0]]], dtype=int32))

Then I get ZeroDivisionError for the center calculation:

def get_centroid_from_contour(contour):
    M = cv2.moments(contour)
    cX = int(M["m10"] / M["m00"])
    cY = int(M["m01"] / M["m00"])
    return (cX, cY)

>>> get_centroid_from_contour(contour)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<stdin>", line 3, in get_centroid_from_contour_
ZeroDivisionError: float division by zero

This I think is related to the fact that somehow opencv thinks the 2x2 squared object has zero area:

>>> cv2.contourArea(contours[0])
0.0

It seems something related to "open objects". Indeed the first contour only contains two points that do not close the polygon, but I have no idea how to fix this. I also tried closing the contour as suggested here but it doesn't work.

Answer 1

Looks like a bug in the implementation of moments(). It appears to not consider how contours are defined in OpenCV.

The issue isn't with your 2x2 connected component, but with the 6x1 one. That one is returned first. Its contour consists of two points.

For the contour of a 1-pixel line, moments() should have still given a non-zero m00 because that is how contours work in OpenCV. Since you say you got a ZeroDivisionError, it must have been zero. That is incorrect for OpenCV's notion of a contour, but correct for every other notion of a polygon/contour.

In OpenCV, a contour describes the polygon that needs to be drawn to reproduce the picture. The line goes ON the outer edge pixels that are still inside the connected component. That is why the contour of your line CC becomes a 2-corner polygon of, strictly speaking, zero area.

OpenCV could have defined contours to be zero-width lines that circumscribe the outer pixels of a connected component. That would have required:

1. the drawing calls in OpenCV to not be as broken and neglected as they are (I'm being blunt but it's true)
2. the conventions for a contour to allow non-integer coordinates

Feel free to submit an issue on OpenCV's github page. The moments() implementation needs a fix.

Addendum: the fix should probably be another flag that distinguishes "contours" from proper polygons, so the mathematically correct behavior isn't lost, but instead made optional. Everyone calls that function on contours, so that should be the default.