[Kwiver-users] detected_object confidence

[Matt Brown]

This is a follow-on from my recent email, but I think it is a new
discussion. A *detected_object*
<https://github.com/Kitware/kwiver/blob/master/vital/types/detected_object.cxx#L42>
has
a confidence associated with it, but it also carries a *detected_object_type
<https://github.com/Kitware/kwiver/blob/master/vital/types/detected_object_type.h#L52>*,
which lets you specify scores for a number of different classification
types. As of now, neither documentation nor implementation enforces
any correlation
between the confidence and the individual class scores. It seems one of the
major use cases for confidence is sorting a *detected_object_set* based on
each *detected_object.confidence*, but this is very confusing in cases
where there are multiple classes associated with the *detect_object_type*.
Detectors with an associated classifier for many classes is becoming more
and more common for us. For example, we might have a faster-rcnn detector
that returns a score for *background*, *person*, *car*, *person-holding-gun*,
etc., and in which case, it isn't clear what the most-useful specification
of the overall detection's confidence would be.

I think the current state of things is a recipe for mistakes and confusion.
One option is to just decide on a convention and document it thoroughly.
Alternatively, we could just remove reference to confidence and require
looking into the detected_object_type scores. Though, I'm not sure I
understand all the use cases that this would break.

Best,
Matt
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://public.kitware.com/pipermail/kwiver-users/attachments/20190313/4a14bcd6/attachment.html>