Contours and shape measurement

Contours turn a mask into countable objects with positions and sizes, the core of classic detection and inspection.

cv2.findContours(mask, mode, method) traces the boundaries of white regions in a binary image and returns (contours, hierarchy) in OpenCV 4 and later. Each contour is an int32 array of points with shape (n, 1, 2).

mode: cv2.RETR_EXTERNAL returns only outer boundaries (usually what you want); RETR_TREE also returns holes, with parent/child links in hierarchy.
method: cv2.CHAIN_APPROX_SIMPLE compresses straight runs to their end points, saving memory; CHAIN_APPROX_NONE keeps every pixel.

With contours you can measure:

cv2.contourArea(c) and cv2.arcLength(c, closed=True)
cv2.boundingRect(c) gives an upright (x, y, w, h); cv2.minAreaRect(c) a rotated box
cv2.moments(c): centroid at (m10/m00, m01/m00)
cv2.approxPolyDP(c, eps, True) simplifies to a polygon; 4 vertices suggests a rectangle
cv2.convexHull, and circularity 4*pi*area/perimeter**2 (1 for a circle)

Filter by area to drop noise, sort by area to find the largest object. cv2.drawContours(img, contours, -1, color, thickness) draws them; index -1 means all.

Feed it a clean binary mask (threshold, Canny, morphology first). Old OpenCV 3 returned three values, which is why older tutorials unpack _, contours, _.

Example

import cv2
import numpy as np

mask = np.zeros((100, 160), np.uint8)
cv2.rectangle(mask, (10, 10), (60, 40), 255, -1)
cv2.circle(mask, (110, 60), 25, 255, -1)
mask[90, 5] = 255                               # one-pixel noise

contours, hierarchy = cv2.findContours(mask, cv2.RETR_EXTERNAL,
                                       cv2.CHAIN_APPROX_SIMPLE)
print(len(contours))                            # 3
objs = [c for c in contours if cv2.contourArea(c) > 20]
for c in sorted(objs, key=cv2.contourArea, reverse=True):
    area = cv2.contourArea(c)
    x, y, w, h = cv2.boundingRect(c)
    m = cv2.moments(c)
    cx, cy = m["m10"] / m["m00"], m["m01"] / m["m00"]
    peri = cv2.arcLength(c, True)
    corners = len(cv2.approxPolyDP(c, 0.02 * peri, True))
    circ = 4 * np.pi * area / peri ** 2
    print(int(area), (x, y, w, h), (round(cx), round(cy)), corners, round(circ, 2))
# 1890 (85, 35, 51, 51) (110, 60) 8 0.87   circle: many corners, round
# 1500 (10, 10, 51, 31) (35, 25) 4 0.74    rectangle: 4 corners

Remember

findContours needs a binary mask and returns (contours, hierarchy)
RETR_EXTERNAL for outer outlines, RETR_TREE for nested holes
Measure with contourArea, boundingRect, moments, approxPolyDP
Filter tiny contours by area to ignore noise

Watch out for

Running findContours on a noisy, un-thresholded grayscale image and getting thousands of tiny contours.

Check yourself

You only want the outlines of separate objects, ignoring holes inside them. Which mode?
How do you compute a contour's centroid x from its moments `m`?

Official docs: docs.opencv.org/4.x/d4/d73/tutorial_py_contours_begin.html

Example

Remember

Watch out for

Check yourself

Builds on

Unlocks