Package in PIP

[f0ma]

Do you have a plan to publish ellipsoid_fit_python in PIP package repository? For example in PIP already available package for ellipse fitting (https://github.com/bdhammel/least-squares-ellipse-fitting).

[aleksandrbazhin]

Hi, I think it needs sufficient refactoring to be published in PIP. In it's current form the repository is more of a example/experiment than a library with a well-designed API. Should plots be included in the library? Should (rather unconventional) data regularization procedures remain? I did not find a suitable answer, so decided to leave it as it is. If you have some thoughts, I am open to discussion.

[f0ma]

Following text is just my habble opinion: Today as I know is no packages in PIP for:

• Ellipsoid fitting by surface points (only for confidence-like ellipsoid like http://kylebarbary.com/nestle/index.html )
• Ellipsoid parametric plotting (but multiple solution available in google/stackoverflow)
• Spherical regularization (???)

All thouse useful methods available in ellipsoid_fit.py File have no dependencies except numpy and optional matplotlib. About API I think ellipsoid_fit and ellipsoid_plot already have a stable API, but data_regularize require optional aggregation method as argument (default np.mean(np.array(points_in_sector), axis=0), also "Other strategy ..." code, may be some problem specific method by user chouse). And tests...I think good idea to create some synthetic dataset to make smoke test. I am ready to prepare test data and unittests and share some real data for making a tests.

[aleksandrbazhin]

About tests - I actually did create synthetic datasets while fixing issues a couple of years ago, just did not commit it. The biggest problem with this repo is - I don't actually use it and created it in a day or so to calibrate an mpu9150 I had in 2015.

Some doubts I have are

• The code for regularization was written in a haste, is naive and may be not correct in mathematical sense, or maybe not called regularization. Is taking mean a regularization at all in a strict sence? I like the idea about passing an aggregation method, though.
• Original matlab code (https://www.mathworks.com/matlabcentral/fileexchange/24693-ellipsoid-fit) has many options for axis aligned ellipsiods which should be also included
• The plotting method relied heavily on the matpltlib option which has been removed (more here https://github.com/aleksandrbazhin/ellipsoid_fit_python/issues/10) - now there is no way to use equal axis scale without downgrading matplotlib. The code for plotting was not written by me at all, it's from https://github.com/minillinim/ellipsoid (which also has some useful functions for ellipsoids - like finding minimum enclosing ellipsoid).
• The output format uses row-vectors instead of column-vectors, which was for my convenience, but is in disagreement with (as far as I remember) both the original matlab version and similar functions in scipy. Though I may be wrong here.

Overlall I agree, that there are some useful functions, and I am happy that someone uses the code. Even disregarding all the concerns above, those useful functions you mentioned are mostly related to each other by their connection to magnetometer calibration, so should the package be also called something like 3d_ellipsoid_calibration?

[aleksandrbazhin]

And another problem with "spherical regularization" is that a center for sphere is a wild guess int the style (min + max) / 2.

Maybe the perfect solution would be join with mentioned minililnim's repository into comprehensive "3d_ellipsoids" package. Or leave everythin as is - it's still googlable.