linear_assignment can now be imported from scipy

sklearn.utils.linear_assignment_ can be replaced by scipy.optimize.linear_sum_assignment now that scipy >= 0.17 is required by scikit-learn.

The sklearn.utils.linear_assignment_ module should be deprecated to be removed in 0.23 (assuming this issue gets fixed before 0.21’s release).

[yiakwy]

@jnhansen @praths007 I am concerned that linear_sum_assignment is not equivalent to linear_assignment which later implements “maximum values” matching strategy not “complete matching” strategy, i.e. in tracking problem maybe an old landmark lost and a new detection coming in. We don’t have to make a complete assignment, just match as more as possible.

I also try to tell others that it is dangerous to use linear_sum_assignment because it implements complete matching, where in the most cases we don’t need.

Instead of using complete matching by KM algorithms, mini cut - max flow might be a good choice.

[thomasjpfan]

Resolved in #13465