Difference between transcription.precision_recall_f1_overlap and multipitch.metrics?

[carlthome]

Considering that transcription.py's symbolic representation is piano rolls with non-quantized pitches, what's the difference between transcription.py and multipitch.py really? I'm extra confused by how transcription.py references the multi-f0 mirex estimation task. Could anyone help clarify why both modules exist? When should which be used?

[craffel]

cc @rabitt @justinsalamon

Well, the difference starts with the fact that their annotation format is quite different. multipitch annotations specify, on a fixed timing grid, "what fundamental frequencies are present at this point in time?" transcription is more like MIDI, specifying individual notes, with start times, end times, and pitches. This means that evaluating them ends up looking a good bit different. I think the confusion about transcription linking to MIREX multi-f0 is that MIREX lumps these two arguably different tasks into one "heading". (@justinsalamon please correct me if I'm wrong and there is in fact an error in the docs). See the annotation format section of the MIREX task description page - there are clearly two different formats/tasks here.

Maybe we can make a PR to clarify this in the documentation?

[justinsalamon]

(@justinsalamon please correct me if I'm wrong and there is in fact an error in the docs).

Nope, that's correct.

[carlthome]

Aha, so multipitch.py is for frame-based pitch estimation while transcription.py is for event-based pitch estimation? Thanks, @craffel!

It could be clarified in the docs, IMHO. Especially pros/cons of the respective formats, with some example use cases. Does it make sense to use multipitch.py for a solo piano? Does it make sense to use transcription.py for a string quartet? I'm confused.

[craffel]

It could be clarified in the docs, IMHO.

I agree. If you or @justinsalamon or @rabitt want to make a PR, it would be welcome.

Especially pros/cons of the respective formats, with some example use cases.

This is outside the scope of mir_eval.

[rabitt]

Another difference - multipitch is intended for time varying pitch
curves/sequences
whereas transcription is events with a single pitch throughout the note.

On Wednesday, September 21, 2016, Colin Raffel [email protected]
wrote:

It could be clarified in the docs, IMHO.

I agree. If you or @justinsalamon https://github.com/justinsalamon or
@rabitt https://github.com/rabitt want to make a PR, it would be
welcome.

Especially pros/cons of the respective formats, with some example use
cases.

This is outside the scope of mir_eval.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#223 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AHypc1Oe0QOIF_REPkV_LWU-IGuNvY2xks5qsbIAgaJpZM4KCzDL
.