rotatedRectangleIntersection does not return all intersection points as expected

[Kroppeb]

System Information

OpenCV python version: 4.7.0
Operating System / Platform: Windows 10
Python version: 3.10.10

Detailed description

During some testing of a system that used rotatedRectangleIntersection and then contourArea to get the intersection area of 2 overlapping rectangles, we got a size of ~3500mm² instead of the expected ~7000mm².

It seems to be related to the fact that one edge of each rectangle are on the same line.

• Rectangle 1: RotatedRect(center=(824.6421183672817, 280.28737007069833), extent=(565.0, 140.0), angle=-177.80506896972656)
• Rectangle 2: RotatedRect(center=(567.3310438828003, 270.42527355719545), extent=(275.0, 50.0), angle=92.19493103027344)

This seems related to #21659, which should be fixed by #21677, but I haven't tested that as it hasn't been released.

Steps to reproduce

import cv2 as cv

res, points = cv.rotatedRectangleIntersection(
    ((824.6421183672817, 280.28737007069833), (565.0, 140.0), -177.80506896972656),
    ((567.3310438828003, 270.42527355719545), (275.0, 50.0), 92.19493103027344),
)

assert res == 1
assert points is not None

points = points.reshape(-1, 2)
area = cv.contourArea(points)
print(area)  # expected ~7000, got ~3500

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)

[dkurt]

Can confirm on 26a7b33 (green are rect1, rect2 and blue color is intersection):

import numpy as np
import cv2 as cv

rect1 = cv.RotatedRect((824.6421183672817, 280.28737007069833), (565.0, 140.0), -177.80506896972656)
rect2 = cv.RotatedRect((567.3310438828003, 270.42527355719545), (275.0, 50.0), 92.19493103027344)
_, inter_pts = cv.rotatedRectangleIntersection(rect1, rect2)

img = np.zeros((512, 1200, 3), dtype=np.uint8)
cv.polylines(img, rect1.points().astype(np.int32).reshape(1, 4, 2), True, (0, 255, 0))
cv.polylines(img, rect2.points().astype(np.int32).reshape(1, 4, 2), True, (0, 255, 0))
cv.polylines(img, inter_pts.astype(np.int32).reshape(1, -1, 2), True, (255, 255, 0))
cv.imwrite("img.png", img)

necessary changes with RotatedRect bindings: https://github.com/dkurt/opencv/tree/py_rotated_rect

[chacha21]

I have tried to use double precision instead of float precision in imgproc.cpp:_rotatedRectangleIntersection()
I confirm that the returned area value becomes 7000.0047071566805