Multiple face detection with PP5 plugin doesn't work about openbr

[jklontz]

@lbestrowden mentioned this issue recently also. Here are my thoughts:

If we would like the option to be able to switch between initializing PittPatt to detect one face vs all faces at run time, then the PP5Context constructor can simply query Globals->enrollAll when deciding how to initialize the PittPatt context. This works because algorithms aren't created until they are first used, giving you the opportunity to specify command line settings before enrollment. Note: this is not the case for anything that subclasses from Initializer, which gets executed before any argument parsing.

As an alternative to changing the interface to allow generating multiple templates, consider storing a list of values in the metadata. Then generalize Expanded in meta.cpp to handle splitting all lists values instead of just points and rects. The crucial step here is that if after the list has been split, if the sublists are of length one then we should store their single element in the expanded template's metadata instead of a list of length one. I think this represents a simpler and more consistent solution for the current code base, though I'd be willing to have a discussion about whether or not we need better support for multi-template project().

Any takers on implementing this?

[caotto]

The overall change to support multi-template project has already been made,
the common use of enroll now passes a series of single template
templatelists through the algorithm (which inherently allows for a variable
number of templates as output). In that sense, just changing PP5Enroll to
implement project(templatelist) would be sufficient. There are some
transforms that haven't been updated (such as Cache), but I'd argue that
just implementing project(templatelist) is simple and consistent with the
current design (if not necessarily implementation) of the codebase.

On Fri, May 17, 2013 at 4:47 PM, jklontz [email protected] wrote:

@lbestrowden https://github.com/lbestrowden mentioned this issue
recently also. Here are my thoughts:

If we would like the option to be able to switch between initializing
PittPatt to detect one face vs all faces at run time, then the PP5Contextconstructor can simply query
Globals->enrollAll when deciding how to initialize the PittPatt context.
This works because algorithms aren't created until they are first used,
giving you the opportunity to specify command line settings before
enrollment. Note: this is not the case for anything that subclasses
from Initializer, which gets executed before any argument parsing.

As an alternative to changing the interface to allow generating multiple
templates, consider storing a list of values in the metadata. Then
generalize Expanded in meta.cpp to handle splitting all lists values
instead of just points and rects. The crucial step here is that if after
the list has been split, if the sublists are of length one then we should
store their single element in the expanded template's metadata instead of a
list of length one. I think this represents a simpler and more consistent
solution for the current code base, though I'd be willing to have a
discussion about whether or not we need better support for multi-template
project().

Any takers on implementing this?

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/54#issuecomment-18085156
.

[jklontz]

Ok, sounds good to me. He who fixes this bug can have final say :)

[jklontz]

Fixed in #55