Jacaard_similarity_accuracy to the objectives

[saeedizadi]

One of widely-used metrics for evaluationg the results of a segmentation task is the Jaccard_similarity_coefficients. It is as simple as diving the intersection region over the union region between two binary maps.

Adding this accuracy metric to the objectives (just like binary_accuracy) can be so helpful.
For more information, please see this Wikipedia page:
https://en.wikipedia.org/wiki/Jaccard_index

[f0k]

One of widely-used metrics for evaluationg the results of a segmentation task

Do you have some references at hand?

Adding this accuracy metric to the objectives (just like binary_accuracy) can be so helpful.

I'd be fine with this. Feel free to submit a PR. There are some things to consider:

• For binary classification, when binarizing the predictions, the Jaccard index is the same as the accuracy.
• One could go for a non-binarized version that computes minimum(A, B).sum() / maximum(A, B).sum(). This would even be differentiable. Probably there's literature for that as well.
• For multi-class problems, this should support both one-hot encoded and integer targets, like categorical_accuracy.
• We should think about whether we want jaccard_index() (jaccard_similarity_coefficient() is a bit long for my taste) or jaccard_distance(). Depends on whether we advertise this as a loss function or evaluation measure.