Problem with ransac predictions.

[bjpietrzak]

I think I have encountered a bug in sklearn.linear_model.RANSACRegressor.

I was using it to simplify OpenCV contours, on the occupancy grid map. Unfortunately, RANSAC was pretty bad at detecting almost straight points perpendicular to the x-axis. I have checked it on a simpler training sample, but still, it's doing the same thing (my RANSAC settings were much more complex than what you can see here).

My code:

ransac = RANSACRegressor()


X,y = np.array([0,5,5,5,5]).reshape(-1,1),np.array([0.5,0,1,2,3]).reshape(-1,1)
ransac.fit(X,y)

inliers = ransac.inlier_mask_

xpred = X[inliers].reshape(-1,1)
ypred = ransac.predict(xpred).reshape(-1,1)

plt.scatter(X,y, color='blue')
plt.scatter(xpred,ypred,color='orange')
plt.show()

Is it my fault or some flaw in the source code?

[glemaitre]

Moving this in discussion for the moment.

[bjpietrzak]

Is there a definition of a linear regression in sklearn that would satisfy
this equation ax+by+c=0?

[glemaitre]

Sorry, I don't understand the question.

[bjpietrzak]

Sorry, is there a linear function in sklearn that uses two parameters to determine the slope of a function, that we could implement into the RANSAC algorithm?

[glemaitre]

Basically, if you provide a matrix X of shape (n_samples, 2), and y of shape (n_samples,) (the target to regress and put it in a LinearRegression you will obtain your linear model with 2 coefficients and an intercept.

[bjpietrzak]

Okay, that fully solves my problem. Thank you!

[bjpietrzak]

Is there a definition of a linear regression in sklearn that would satisfy this equation ax+by+c=0? W dniu pon., 5.09.2022 o 15:21 Guillaume Lemaitre ***@***.***> napisał(a):

…

So I just plotted your data: [image: image] <https://user-images.githubusercontent.com/7454015/188459027-4ecb3482-f08d-4d72-a8e7-c68d9fe18736.png> I assume that you intuitively would like to consider the first point as an outlier. However since you are using linear regression of the form: y = coef_ @ X + intercept_, there is no perfect model that draws a vertical line that links y and X. So I am not sure what you are actually expect from running a RANSAC on those data. — Reply to this email directly, view it on GitHub <#24273 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ASBRPSCOACOEYE4ATDXIDSLV4XXVLANCNFSM6AAAAAAQE7TN5M> . You are receiving this because you authored the thread.Message ID: ***@***.*** com>