boundingRect returns incorrect result for mask with some image/shape combinations

[damonmaria]

System Information

opencv-python-headless: 4.8.0.76
OS: MacOS 13.5.1
Python: 3.11.4

Detailed description

Using cv.boundingRect to find the bounding rectangle of a uint8 mask image (all 0 or 255) works in the vast majority of situations for me but returns the wrong values for some specific combinations.

Steps to reproduce

import numpy as np
import cv2 as cv

for img_w in range(3, 20):
    for img_h in range(1, 15):
        img = np.zeros((img_h, img_w), np.uint8)
        img[-1, :3] = 255
        x, y, w, h = cv.boundingRect(img)
        if w != 3:
            print(f"{img_w}x{img_h}: {w=}")

boundingRect should return w as 3 for all these combinations of img_w and img_h. But several combinations return w as 1. The other values (x, y, h) seem to be correct.

3x4: w=1
3x8: w=1
3x12: w=1
5x2: w=1
5x6: w=1
5x10: w=1
5x14: w=1
7x4: w=1
7x8: w=1
7x12: w=1
9x2: w=1
9x6: w=1
9x10: w=1
9x14: w=1
11x4: w=1
11x8: w=1
11x12: w=1
13x2: w=1
13x6: w=1
13x10: w=1
13x14: w=1
15x4: w=1
15x8: w=1
15x12: w=1
17x2: w=1
17x6: w=1
17x10: w=1
17x14: w=1
19x4: w=1
19x8: w=1
19x12: w=1

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues, forum.opencv.org, Stack Overflow, etc and have not found any solution
• I updated to the latest OpenCV version and the issue is still there
• There is reproducer code and related data files (videos, images, onnx, etc)

[damonmaria]

is because the white pixel at the bottom of the image is not connected to any other white pixels

@iknothing In all of those images generated in the code I posted there are 3 white pixels in a row, all connected, extending right from the bottom right corner. There is never a single pixel by itself.

As a simplified example:

img = np.array([
    [0, 0, 0, 0, 0],
    [1, 1, 1, 0, 0],
], np.uint8)
print(cv.boundingRect(img))

Prints: (0, 1, 1, 1), so a width of 1.

Whereas:

img = np.array([
    [0, 0, 0, 0, 0],
    [0, 0, 0, 0, 0],
    [1, 1, 1, 0, 0],
], np.uint8)
print(cv.boundingRect(img))

Prints: (0, 2, 3, 1), which correctly has a width of 3.

[crackwitz]

I've replicated the issue with images having shape (h,w,1). The glue logic that converts to cv::Mat seems to be unaffected.

Here's a scatterplot of those (w,h) pairs for which boundingRect() seems to fail. Perhaps the pattern means anything.

[crackwitz]

Tests in C++ appear to produce contradictory results:

• https://forum.opencv.org/t/doubt-in-bounding-rectangle-algorithm/14495/12
• https://forum.opencv.org/t/doubt-in-bounding-rectangle-algorithm/14495/13