[Digikam-devel] Re: Face detection

[Sven Burmeister]

Am Donnerstag, 4. November 2010, 15:31:04 schrieb Marcel Wiesweg:
> > If I select a tag within "People", e.g. People > Person 1 it shows me all
> > the phtos associated to that person. Having that view I start the
> > scanning+merge for another folder. Faces found are now added to the
> > People
> > 
> > > Person 1 view instead of unknown or the proper tag, i.e. People >
> > > Person
> > 
> > 2. In fact the Person 2 does not show any new pictures found althought
> > there are quite a few so it should have found at least one that matches
> > Person 2.
> > 
> > So to me it seems that found faces are always added to the tag that one
> > currently views while scanning.
> 
> Does this assignment remain after switching the view or closing digikam, so
> it's really the wrong tag assigned?

Yes

> Is this reproducible also for other tags, so it is not caused by a broken
> face recognition which assigns Person 1 to almost anything?

Indeed they are always assigned to Person 1. And this happens everytime I 
"scan and merge" the same folder although I mark those pictures as "not this 
person", i.e. click on the "-" in the mouse-over GUI.

I think it really is the first tag I added.

And those faces confused are already quite well trained, i.e. > 50 pitures 
confirmed for each.

> > Another thing I do not understand is of what purpose it is to have some
> > checkboxes in the "pick albums for scanning" in a brighter colour than
> > others.
> 
> I can confirm this, it's probably a subtle gradient applied by the oxygen
> style. Going towards the bottom, the color becomes darker.

Here the first checkbox(es) are dark as well.

> > Also it seems that the list scrolls automatically if one moves the
> > mouse cursor to its bottom but not when moving it to its top.
> 
> Yes. If you hover an item, you "select" it. Qt always shows the complete
> currently selected item. In the top, there is never a partially visible
> item, it's always scrolled to an exact item border. At the bottom, there
> is often a partially visible item, which becomes selected on mouse over,
> is shown fully, and so on.
> If this is undesirable, we can disable autoscrolling.

As far as I can judge there is no partly visible item and it happens always 
not just sometimes, i.e. for some items. Unless one scrolls to the bottom of 
the list it is not possible to select the last item in the list because as 
soon as one touches it, the scrolling starts and not just for that possibly 
partly shown item but until the list ends.

> > Lets say one confirmed Person 1 for a picture where actually Person 2 is
> > shown. IMO there is no obvious (easy enough) way to correct that mistake.
> 
> I'm unsure atm if training can be undone. Need to ask Alex.
> 
> > If one views People > Person 1 and removes the Person 1 tag from an image
> > the image is not removed from the view. If one opens the picture one sees
> > that it still has the face marked as Person 1. Clicking on the "Person 1"
> > tag does not give any possibility to change or remove that tag.
> 
> In preview, you'll need to click on the label with the name to open that
> one for editing again.

Nothing happens if I click on the label in the preview, neither right nor 
left-click does anything.

> In the icon view, there's still the Reject overlay
> to remove, though you're right, there should be an overlay for editing.

Only to remove the whole picture i.e. train the facerecognition that this is 
not a face at all. That seems wrong.

The picture whose Person 1 tag I removed and which still shows up in the 
Person 1 tag view does not show any mouse-over GUI anymore.

> > For pictures in the "unknown" tag if one clicks on "Who is this" there is
> > no way to select an already existing tag but one has to type the name
> > again.
> 
> You'd prefer the combo box like in the preview mode?

In preview mode, i.e. clicking on a picture when viewing all pictures with 
People > Person 1 tag, clicking on "add face" does not do anything. So I 
cannot answer this yet.

> Note that the last typed text remains, which can be both convenient or
> annoying.

Yes, I thought about this too. It is handy if one uses it to assign the same 
name to all pictures where it fits, then changes the name and does so again. 
Yet it feels like a feature which is not meant to work that way but was 
introduced by coincidence. I think a bigger list of alread available tags 
would be most convenient, including some picture of those tags, i.e. something 
like:

Possible matches:
picture of a person 1 | tag-name
picture of a person 2 | tag-name
picture of a person 3 | tag-name

One could actually even just display the photos as possible matches since this 
whole thing is about face recognition and not name recognition. :)

Possible matches:
picture of tag person 1 | picture of tag person 2.

Of course for this digikam would have to be able to at least tell whether some 
faces look similar, i.e. group as mentioned above.

This gets me to another point. Would it be possible to not only have unknown 
pictures but groups of unknown, i.e. digikam should be able to guess which 
faces might look the same although it does not know their name.

And also, what happens if a user scans the whole collection and ends up with 
thousands of pictures in unknown? He would have to assign tags to each of them 
and digikam would not try automatically to guess faces again after it got some 
training. So the user would have to know about the "recognise again" feature 
in the scanning GUI. I think there should be some smartness in digikam that at 
least asks after x pictures training for a tag whether it should try to find 
all pictures for that person.

Also, since digikam only uses tags, is it possible to add more to a face, i.e. 
some data about that person, e.g. a photo (as icon for that tag), which 
familiy/group of persons (other persons/tags) he belongs to? 

> > If one starts typing the already existing tag suddenly appears in a
> > drop-down menu. Moving the mouse on the drop-down menu and out of it
> > closes the drop- down menu which should not happen.
> 
> It happens when you touch a different picture with the mouse. That's the
> idea of a mouse-hover overlay... perhaps we need some sort of locking for
> the cases when the overlay opens a menu.

If the user typed something the pop-up should not close until the user 
confirms or erases what he typed.

> > If one wants to add/confirm etc. tags while digikam still scans the view
> > is refreshed constantly and makes working on the pictures impossible. I
> > think this is related to the bug that restarts videos when one plays
> > them from within a folder that new pictures are added to.
> 
> It's a similar problem, but probably not the same solution. But it has the
> same solution as the mouse-over-drop-down-menu problem above.
> When images are inserted before the current picture, its position will
> change. We can think of a solution with a "persistent editor" if the line
> edit currently has the focus, but it's not too easy to implement
> (currently, the overlays take the easy way of hiding at any indication of
> change.)

It also happens if a picture is added below the picture one is currently 
typing a name for.

Sven