Issues with EllipseModel

[rk-exxec]

Description

Hey, I have issues with the EllipseMopdel estimate function.
It produces a completely wrong ellipse, what am I doing wrong.
This is what it looks like:

I am detecting a contour with opencv (blue line) and then feeding that contour to the EllipseModel.

Way to reproduce

import math
import cv2
import numpy as np
from skimage.measure import EllipseModel
def evaluate_droplet(img, y_base):
    crop_img = img[:y_base,:]
    shape = img.shape
    height = shape[0]
    width = shape[1]

    bw_edges = cv2.Canny(crop_img, 76, 179)

    contours, hierarchy = cv2.findContours(bw_edges, cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE)

    if len(contours) == 0:
        raise ValueError('No contours found!')

    edge = max(contours, key=cv2.contourArea)

    img = cv2.drawContours(img,edge,-1,(255,0,0),2)
    points = edge.reshape(-1,2)
    ell = EllipseModel()
    if not ell.estimate(points): raise RuntimeError('Couldn\'t fit ellipse')
    x0, y0, a, b, phi = ell.params

    img = cv2.ellipse(img, (int(round(x0)),int(round(y0))), (int(round(a)),int(round(b))), int(round(math.degrees(phi))), 0, 360, (255,0,255), thickness=1, lineType=cv2.LINE_AA)

if __name__ == "__main__":
    im = cv2.imread('untitled1.png')
    try:
        drp = evaluate_droplet(im, 250)
    except Exception as ex:
        print(ex)
    cv2.imshow('Test',im)
    cv2.waitKey(0)

Version information

>>> import sys; print(sys.version)
3.8.6 (tags/v3.8.6:db45529, Sep 23 2020, 15:52:53) [MSC v.1927 64 bit (AMD64)]
>>> import platform; print(platform.platform())
Windows-10-10.0.18362-SP0
>>> import skimage; print("scikit-image version: {}".format(skimage.__version__))
scikit-image version: 0.17.2
>>> import numpy; print("numpy version: {}".format(numpy.__version__))
numpy version: 1.19.2

Original image:

[jni]

I gotta say I can't figure this one out after a bit of playing with it. Anyone in @scikit-image/core have experience with ellipse estimation and have time to play with this?

[mark-boer]

This looks interesting, I have some experience with fitting geometric shapes (mostly 3D tho), I'll have a look, maybe I can figure it out.

[mark-boer]

Found it! It is an integer overflow in

scikit-image/skimage/measure/fit.py

Line 426 in 9d502e2

S1 = D1.T @ D1

Just convert your input to floats, points = points.astype(float). That should fix your issue.

Also, be careful with using opencv and scikit-image, they have slightly different definitions of the order of coordinates 😉

@maintainers we should probably convert the input of CircleModelestimate.estimate and EllipseModel.estimate to floats, because they are pretty much guaranteed to overflow.

Happy coding!

[grlee77]

@maintainers we should probably convert the input of CircleModelestimate.estimate and EllipseModel.estimate to floats, because they are pretty much guaranteed to overflow.

Thanks for looking into this! Indeed, it seems like casting data to float should be done within these functions. Are you interested in making a PR for this? (no worries if you want us to handle it, just let us know)

[grlee77]

p.s. I don't think maintainers pings anyone here? Most likely @scikit-image/core was what you wanted, but in general we try to monitor new issues, so usually it won't be necessary to specifically ping the full team unless you think something is falling through the cracks and needs special attention.

[mark-boer]

Thanks for looking into this! Indeed, it seems like casting data to float should be done within these functions. Are you interested in making a PR for this? (no worries if you want us to handle it, just let us know)

Sure, I'll create one this weekend.

[rk-exxec]

Thank you mark!

[rfezzani]

Closing this since #5179 is merged. Thank you again @mark-boer 😉