Whats your workflow for person keywording

[ben]

Hello everyone,

i am wondering what's your workflow regading keywording for the new "person" feature.

I used to add hierarchical keywords for all (interesting persons) like "myperson | name".
I think some of you had similar keywords schemes.
The new face recognition allows me to automatically add keywords to recognized persons, which is great.
Unfortunately, i haven't found the right workflow yet.

Are you going to use manual and automatic keywords in parallel from now on?
Or do you use only either one?

Mario suggested to use different manual and automatic keywords.
For me, it would be perfect if manual keywords and automatic keywords (from face recognition) were somehow synchronised.
Using the same manual and automatic keywords has some drawbacks in some situations, though (see thread https://www.photools.com/community/index.php?topic=9869.0).
Using different keywords would mean, there are some images from "John" only visible in the "people" view and some only in the "categories" view, which i don't like.

So, it would be great to hear how you are working with the new feature.

Thanks for ideas
Ben

[Jingo]

Well - it is still early and I'm rethinking my scheme at the moment based upon the future "remove fingerprint" options Mario laid out in another post.  I think the biggest issue is for folks trying to use the AI on a very large age gap in the database.  Once the fingerprint removal option is enabled, I think my workflow will go something like this:

a) use the AI on a person at a certain age (say baby to 2 years old)... train the AI on that and add the images to the PERSON (with hierarchical keywords already assigned)
b) use the fingerprint removal tool to take out those faces but still leave the images and face annotations in place.
c) use the AI on the same person for the next age rage (say 3-5)... rinse and repeat until we reach the current age.

At that point - the face will not change that much for awhile and we can probably leave things in place moving forward for new images to automatically get added to the person and for the keywords to get added as well.

For me - the face recognition is more useful as a tool to quickly and automatically add keywords to the right person rather than trying to find specific faces WITHIN the image... so this should work well for me and my existing database and new images.

[ben]

For me - the face recognition is more useful as a tool to quickly and automatically add keywords to the right person rather than trying to find specific faces WITHIN the image... so this should work well for me and my existing database and new images.

Interesting idea.
So, would you also not use the new People view?
Instead you keep using the category view with your own "persons" tree?

So basically you would let the face recognition add keywords.
Then in the next step you might add missing keywords for not recognised persons?

BEn

[Jingo]

For me - the face recognition is more useful as a tool to quickly and automatically add keywords to the right person rather than trying to find specific faces WITHIN the image... so this should work well for me and my existing database and new images.

Interesting idea.
So, would you also not use the new People view?
Instead you keep using the category view with your own "persons" tree?

So basically you would let the face recognition add keywords.
Then in the next step you might add missing keywords for not recognised persons?

BEn

Correct... if "People" remain linked only to "face detection" and actual bounding boxes / Annotations - then yes... I might not use the People View.  If People were extended to allow a connection or link to keywords as well.. then I would most certainly use it! 

For me, the ideal scenario is to use the People function and AI to identify folks moving forward... but be able to link existing images to the Person via the linked keyword(s).  So, an @Keyword of People->Family->Beth  would link to the person "BETH" even if that person wasn't setup via the AI (ie: older photos).   Not sure if Mario would agree to this type of implementation of the Person function... but to me - AI is only a part of the feature and future advancements to the Entity functionality would then become available to me without having to go back and fingerprint many thousands of already keyworded images.

[Mario]

A person is an object in the database.
You can create any number of persons you want.

But persons don't exist in thin air.
To link a person with a file, you have to link it to a face in that file. Even if you link it just to a virtual face you create as an annotation.

[Jingo]

A person is an object in the database.
You can create any number of persons you want.

But persons don't exist in thin air.
To link a person with a file, you have to link it to a face in that file. Even if you link it just to a virtual face you create as an annotation.

That is an interesting idea... and perhaps that is something I might envision doing... just create a "dummy" face in the image that would create the link... hmmmmmmmm - will think more about that tomorrow.

[Mario]

But if you have to create a dummy face annotation - is there no person in the image? If not, why link a person with it...?

[plastikman]

I think some people over think this feature.

From my experience I manually tag most pictures with a person keyword, then use that as a filter and run the face recognition within it. This creates perfect training. Once people are properly trained, all future photos will be properly ID-ed.

Of course it depends on the type of photography you do but for me it is mostly personal family and travel photography and which means this system works and I find the other way around it just creates more work since the AI misses faces so you have to manually go through everything anyway.

In my opinion, the easiest way to do this is use either timeline or folder view and just go through day-by-day or event-by-event. Use a grid/zoomed out thumbnail view and just select all pictures with a person, manually add keyword, and move to next person. Takes a few minutes of work. Then go to filter/category view with person X and show face annotations and just skim through them and confirm.

One shortcut I found is that you can copy a face annotation box. So just copy a box with person X, and then go through all the missed face annotations for the person and just copy the box, drag it in the right place and move on to next image.

Other thing I use in viewer mode is Ctrl + A to select all face annotations and then C to confirm (when the unconfirmed suggestions are the right person).

[Jingo]

But if you have to create a dummy face annotation - is there no person in the image? If not, why link a person with it...?

There would be the actual person in the image.  This is to handle the case where I have people who are already keyworded as babies up through adults...  I don't truly need to face identify the younger folks as I only want to use the AI moving forward.  HOWEVER, I do wish to use the People feature on that person.  Since the system doesn't just link all images of the Person by keyword and requires a face - the only way I can use the people feature for the older images would be to perform the AI.  Of course, that creates a fingerprint with older images and it messes up AI for the future faces.

So, if I can somehow just create a "dummy" face box on the image (not even on the actual face) for that person, then they will all appear in the People View.  Not ideal.. but not sure how else to link thousands of older keyworded images to an existing person without "polluting" the AI.  Also, it is lots of work to manually set approve, setup thousands of old images...

I thought I could create an app to just pull up a group of images by keyword and if they don't have a person annotation, simply add a tiny one to the image.. but there are no documented annotation endpoints and the viewer is of course not written as an app so I can't borrow code.   

[Jingo]

I think some people over think this feature.

From my experience I manually tag most pictures with a person keyword, then use that as a filter and run the face recognition within it. This creates perfect training. Once people are properly trained, all future photos will be properly ID-ed.

Of course it depends on the type of photography you do but for me it is mostly personal family and travel photography and which means this system works and I find the other way around it just creates more work since the AI misses faces so you have to manually go through everything anyway.

In my opinion, the easiest way to do this is use either timeline or folder view and just go through day-by-day or event-by-event. Use a grid/zoomed out thumbnail view and just select all pictures with a person, manually add keyword, and move to next person. Takes a few minutes of work. Then go to filter/category view with person X and show face annotations and just skim through them and confirm.

One shortcut I found is that you can copy a face annotation box. So just copy a box with person X, and then go through all the missed face annotations for the person and just copy the box, drag it in the right place and move on to next image.

Other thing I use in viewer mode is Ctrl + A to select all face annotations and then C to confirm (when the unconfirmed suggestions are the right person).

My experience - this only works well if the person is relatively the same age.  I have thousands of images for people that range from birth all the way to age 20.... the AI works great within an age "bracket" but then confuses the person with other family members as they get older. 

Also - having to go through and manually correct or verify thousands of already tagged images is a LOT of work currently as setup in the Viewer...  down the road, this might be easier with a "face thumbnail" method of confirm/correct... where just the face for the image is shown and you can quickly confirm or correct them en masse.

[ubacher]

down the road, this might be easier with a "face thumbnail" method of confirm/correct... where just the face for the image is shown and you can quickly confirm or correct them en masse.

I think we don't have a feature request for this - do we?
I have suspended my work with face recognition until we have this feature. ( I.e. I consider it so essential that it sure to come).

[Jingo]

down the road, this might be easier with a "face thumbnail" method of confirm/correct... where just the face for the image is shown and you can quickly confirm or correct them en masse.

I think we don't have a feature request for this - do we?
I have suspended my work with face recognition until we have this feature. ( I.e. I consider it so essential that it sure to come).

Correct.. not yet but it has been discussed in other threads so we are hopeful something will come that makes reviewing/approving the AI a bit easier than an image by image review in the Viewer.

[Mario]

Developing something like this will take a while. Not for the next release.

[Jingo]

Developing something like this will take a while. Not for the next release.

Yes.. totally understandable..