Confirming Faces: Batch method

[Jingo]

This is a feature request for handling a batch method of auto-assigning a person to unconfirmed face annotations.

Let's say I have a portrait session.  The images are all very similar - it is the same person taken across 200 images.  When I run the face recognition, IM is finding faces - but there are a mix of correct and incorrect unconfirmed faces assigned.  Since there are incorrect unconfirmed faces, I cannot use the context menu option to confirm in batch.  So, this request would be to create a few methods of confirming a single person to the unconfirmed images. 

I'm thinking a few methods:
  1 - Right click option that would allow you to: Assign Person to selected images...   - then, the Assign Person box would open and you would select or add a new person to all the selected images.
  2 - Drag/Drop the selection onto the Person in the Person View window (Thorston's Idea from his post)

Along with this (as an aside), it would be nice if there was a place to SEE the unconfirmed data associated with the annotations in bulk outside the viewer.  This way, we could confirm only a single face is selected and/or what the unconfirmed selections are.  If this data is available via a GET endpoint, I'm thinking I can create an APP... I'll need to investigate how the annotations are store for this in the DB.

[graham1]

1 - Right click option that would allow you to: Assign Person to selected images...   - then, the Assign Person box would open and you would select or add a new person to all the selected images.
  2 - Drag/Drop the selection onto the Person in the Person View window (Thorston's Idea from his post)

The Like button does not see to be working, so I would like to add my +1 to this.

Graham

[Mario]

The Like button does not see to be working, so I would like to add my +1 to this.

Pls read the "Likes are broken" sticky topic at the very top of this board.

[hluxem]

+1

[Ron_S]

+1

[jch2103]

+1

[sinus]

+1

[thrinn]

+1

[mastodon]

Would it be better to do the batch confirming like picasa do? I only see the faces, that need to confirm as "John", and if I click means confirmed, and after toggleing a switch, I can uncomfirm the rest. And one could select lot of face thuimbnails in a row to confirm.

[Mario]

Would it be better to do the batch confirming like picasa do? I only see the faces, that need to confirm as "John", and if I click means confirmed, and after toggleing a switch, I can uncomfirm the rest. And one could select lot of face thuimbnails in a row to confirm.

That's easy to do:
Click on John in the People View. This shows all images assigned to John (confirmed or not).
Use the People Filter to see only unconfirmed.
..

[ben]

Would it be better to do the batch confirming like picasa do? I only see the faces [...]

That's easy to do:
Click on John in the People View. This shows all images assigned to John (confirmed or not).
Use the People Filter to see only unconfirmed.

The "show only the face area" in Picasa is perfect.
The difference to the people view is, that you can tell immediately if the right face has been tagged.
Think of a group image. Even if you click on John and filter for unconfirmed you cannot be sure if the right face has been detected for John. All you can be sure of is that John is on the picture

[Mario]

I have never used Picasa - except for looking at the metadata mess it crated and how to make it export face annotations to XMP.

I guess you want the file window to change somehow (?) to not show the image itself but the face rectangles with content? What the Viewer and the Quick View Panel already do?

[ubacher]

How to confirm people is certainly an issue. I am not sure yet what to suggest- haven't played with it enough.

I did find the way Picasa does it well done. Certainly having to open the viewer for confirming is cumbersome.
I suggest: Let's collect ideas as to how best do this before we settle on a method.

[mastodon]

Mario, could you please post a sreenchot of What the Viewer and the Quick View Panel already do. If there is then better.

[ben]

I have never used Picasa - except for looking at the metadata mess it crated and how to make it export face annotations to XMP.

Well, that's why we all use iMatch, right!

I guess you want the file window to change somehow (?)

Either the file window or an app window that pops up, or whatever works.

to not show the image itself but the face rectangles with content? What the Viewer and the Quick View Panel already do?

The image would be cropped to roughly the size of the face that is to be confirmed
Then we can scroll through many images very quickly and we are sure that the correct faces have been selected, e.g. in a group image

[hluxem]

When I converted my database a message popped up saying I have ~ 42K face annotations and 665 persons out of 240K files. I started with face tagging ~ 2015, pictures added to the database before that have no face tags.

I do try to make things as efficient as possible and used mostly 2 methods for that. When I had a lot of pictures with many different faces I used the last standalone version of LR to recognize and create the face tags. For most newer pictures with just some family members I used copy and past mostly. Not enough to go through the hassle of synchronizing with LR.

I used Picasa only a little and I think it worked similar to LR. LR shows thumbnails of the faces detected, not images. That could be 200 faces in one window. Instead of selecting and confirming one face at a time in the viewer, you can select and confirm multiple faces. Hundreds of faces in minutes is easily done that way. I did run face recognition on about 300 pictures last night. The face recognition works very well, the suggestions are good, but not perfect. With mostly family members I'm not surprised when my dad and my brother gets tagged wrongly. I also wonder if having that many persons makes it harder for the system to create good fingerprints. I think having a batch method with thumbnails could cut the time for confirming up to a factor of ten. 

If users start to run face recognition on their entire database, fast and efficient confirming of faces will be important, no mater how good the face recognition is.
As an added benefit, having thumbnails of the face allows you to easily review face annotations. Not important for working with faces, but it's fun to see how faces change over time.

[ubacher]

Picasa shows you  (all at once as thumbnails) faces (only) to confirm as soon as you identify a person.
Once you confirm it searches again, based on the updated face-fingerprint and then it presents you
more faces to confirm. Several iterations can follow. I assume this process cuts down on faces being wrongly
identified based on the initial small sample. This then makes the confirmation process faster because fewer
miss-identifications occur.

Does Imatch re-evaluate found (unconfirmed) faces once it has gotten more confirmed ones i.e. once it has an improved
face-print?

[Mario]

Does Imatch re-evaluate found (unconfirmed) faces once it has gotten more confirmed ones i.e. once it has an improved
face-print?

Please see https://www.photools.com/help/imatch/#people.htm

[axel.hennig]

+1

[Mario]

Picasa shows you  (all at once as thumbnails) faces (only) to confirm as soon as you identify a person.

Please all remember how many programmers Google has (150,000 employees) and how many people Mario is... 

[Jingo]

Picasa shows you  (all at once as thumbnails) faces (only) to confirm as soon as you identify a person.

Please all remember how many programmers Google has (150,000 employees) and how many people Mario is... 

And we appreciate your efforts for sure!  I'm always happy to help code apps when (if) endpoints become available... Not the same things as a staff of programmers... but another group of free labor to assist!

[ben]

Picasa shows you  (all at once as thumbnails) faces (only) to confirm as soon as you identify a person.

Please all remember how many programmers Google has (150,000 employees) and how many people Mario is... 

1 Mario equals at least 1000 standard programmers, so the equation is not that bad 

[Mario]

1 Mario equals at least 1000 standard programmers, so the equation is not that bad 

Don't be fooled. Google hires the smarted guys and gals available, from everywhere.

[mastodon]

This is a feature request, and we all know, that you have limited time. But this batch face confirmation would be a very nice one. Picasa has a good start on this, and we like IMatch to rule this field too.

[PaulS]

+1 to the Picasa method.  Works very nicely to do a large number of photos at once with high confidence the face is being tagged correctly.

[Mario]

+1 to the Picasa method.  Works very nicely to do a large number of photos at once with high confidence the face is being tagged correctly.

You can use Picasa. Let it write XMP. Then import the images into IMatch 2020.
This would give you access to the Picasa Batch method immediately.

[ubacher]

When adding face detection to a large existing database it is very frustrating to get a large number of false matches.

I assume this is why Picasa first sets the match criteria very high and as the persons face-print gets more
confident it reduces the number of false matches as it re-scans. Avoids frustration.

Now I would think Imatch re-evaluates the found and unconfirmed faces with each face confirmed. My gut feel, from the
testing I did so far is that there is something else wrong which produces a lot of miss-matches. Many of the miss-matches
I encounter seem so wrong that I came to this conclusion. And the miss-matches don't seem to correlate with the number of
faces the subject face-print is made up of (which I could understand).

My conclusion: Having a way to confirm faces by only seeing the face that needs to be confirmed (and not the whole image) is definitely
desired.
The iterative method of confirming faces my not be needed if the problem with finding too many false matches can be found and corrected.

+1

[Mario]

The iterative method of confirming faces my not be needed if the problem with finding too many false matches can be found and corrected.

This is the first time something like this is reported.
False matches are normal. And also some obscure effects - that's just how AIs work. But that's less than 5% of all cases, together.

I can only comment on what you are stating by having your actual database in my lab. Or at least a representative sample of the images you process and which produce the strange results you are stating.
There are no parameters for the AI, no bugs to fix (unless, for some obscure reason) IMatch fails to extract face data in the correct way from your images...
I have tested over 100,000 files from all kinds of sources, decades and motives, lightning conditions, genders, ages, skin colors etc.
And the AI model was trained on over 300,000 faces.

[PaulS]

You can use Picasa. Let it write XMP. Then import the images into IMatch 2020.
This would give you access to the Picasa Batch method immediately.

Yes, this is very close to what I do right now.  Since Picasa messes up metadata when it writes faces, I instead use EXIFTool to copy only the face data from the Picasa modified file back to the original in IMatch and then rescan the metadata to import the faces.

In my short experience with IMatch 2020, I find it more user friendly than Picasa for people and faces, other than batch confirmation of faces.  Face recognition seems to be good, although I so far have limited experience with IMatch 2020.

So I really appreciate what you have accomplished with IMatch 2020 and am close to being able to dump Picasa for good.

[Tveloso]

I would like to propose a hypothetical function for this, based upon this image from the IMatch Help:

   

The Panel on the left is the People View.

So is the one at the right, but...let's imagine that the Panel on the right is actually presenting us with all Faces for the Person selected in the People View at the left - either only the Confirmed Faces for this person (to allow for a quick review, to check for assignment errors), or only the Unconfirmed Faces (to allow for the Batch Face Confirmation). 

That second panel could provide just "Confirm" and "Not this Person" commands (via either a right-click menu, or separate buttons at the top of the panel).  We could then Ctrl-Click and/or Shift-Click a range of faces, and use one of those two commands.  It might also be nice to be able to click on a single face in that "all faces panel", and press Ctrl-G to go to the image (in the Media and Folders view) where that face came from (so we can open it in the Viewer and make adjustments to the placement of the Annotation).

This functionality would be in addition to (not instead of) the functionality provided in the Viewer.  Perhaps opening that second "all faces panel" could be done via the additional functions button in People View, and maybe it could replace the screen area that's occupied by the People View itself (rather than opening a separate window, or additiona panel).

[Mario]

This has already been discussed to death now and as I said in several posts, I will provide a new feature in one of the next releases. I'm already ahead.
Currently I'm concentrating on bugs reported, on the next release and on dealing with the Corona virus spreading and affecting everything here...

[Frank]

+1

[Tveloso]

Thank you so much Mario.

I hope that you and your family are staying safe, and are not directly affected by the Corona Virus.  Of course, as you said, it's impacting us in so many other ways.

[plastikman]

Would it be better to do the batch confirming like picasa do? I only see the faces [...]

That's easy to do:
Click on John in the People View. This shows all images assigned to John (confirmed or not).
Use the People Filter to see only unconfirmed.

The "show only the face area" in Picasa is perfect.
The difference to the people view is, that you can tell immediately if the right face has been tagged.
Think of a group image. Even if you click on John and filter for unconfirmed you cannot be sure if the right face has been detected for John. All you can be sure of is that John is on the picture

This is the problem. In the people view, you see all images with the person in it. However, often there are multiple people in that photo. And occasionally, IMatch tags the wrong person or the same person twice. There is no way to discern this in the people view.

This is why a zoomed in view of the actual identified person(s) is essential: it will allow us too quickly take out the faulty tags. The new trained faces pop-up window already implemented it right. I believe the people/person view should have a toggle for this type of view.