LR6 Facial Recognition - Experiences

Status
Not open for further replies.

Linwood Ferguson

Linwood Ferguson
Lightroom Guru
Joined
Jan 18, 2009
Messages
2,587
Location
Fort Myers, FL
Lightroom Experience
Advanced
Lightroom Version
Classic
I've spent the last several hours playing with facial recognition. I like it a lot, with a few "buts". I thought it may be useful to share experiences here, so please do if you agree.

I did not process my whole catalog (46k photos) as I did not want to wait that long. Instead I would select a year at a time and process them.

The good:

- For the most part once I got use to it, the interface was pretty nice, easy to use quickly.

- I did not find any cases where the "similar" photos were not the same person.

- I found many, many cases where the guess it makes is incorrect, but the process doesn't confirm those automatically, so that's good, and very easy to mass-select and correct.

- The tagging display on loupe is pretty nice as well. A bit difficult in crowd shots, but not bad.

The bad:

- VERY often I went to type a name and found I wasn't fully in the entry mode -- each letter I typed then did something (i.e. shortcuts). Can make a real mess that way, very fast. They REALLY need a more firm way to lock you into text entry when doing People things. I also think there's a delay there -- type too fast even when you did everything right and it responds as a shortcut as opposed to text entry.

- I went through a couple years thinking this is easy and fast and... then I looked more carefully. I found it missed a LOT of faces -- probably almost as many as it found. In particular it missed faces not face-on, i.e. profiles, sideways, sometimes whole-frame faces. I went through a few shots at parties, sports games, etc. which it had completed, and found maybe 30% of the images had no faces identified when there were clear faces, and about 50% of the images had faces found, but not all of them. Here's an example -- it found the face on the left, not the one on the right, despite being (to a human) awfully clear and similarly visible. So if you plan to be fairly complete in your indexing, you are going to have to look at each photo.

attachment.php


This latter aspect is, to me, by far the most disappointing. It's one thing to go through and correct bad guesses, or fill in blanks, but to have to view each photo to ensure it did not miss photos is very tedious.
 

Attachments

  • faces.jpg
    faces.jpg
    8.1 KB · Views: 572
Last edited by a moderator:
"The bad:

- VERY often I went to type a name and found I wasn't fully in the entry mode -- each letter I typed then did something (i.e. shortcuts). Can make a real mess that way, very fast. They REALLY need a more firm way to lock you into text entry when doing People things. I also think there's a delay there -- type too fast even when you did everything right and it responds as a shortcut as opposed to text entry."


Have to agree I kept getting the short-cuts as opposed to entry mode.

Also noticed it picks up objects such as cakes, stomachs etc weird!

More work required me thinks.
 

Also noticed it picks up objects such as cakes, stomachs etc weird!

But entertaining (at least for a while), it picked up Sue (the dinosaur skeleton in the Chicago museum) very nicely, identifying her face).

I'm more concerned how often it is missing faces entirely. False positives are easy to exclude, false negatives there is no way to find but going back through images individually. The more I use it the more I see that happening, and not tiny or indistinct images, but big faces clearly visible (like my example above).

That there is no real way to re-scan to add is going to be really annoying. Imagine when LR 6.0.1 comes out with "dramatically improved face recognition" and 40,000 images have already been scanned and no way to say "try again and add what you missed". I hope they consider this issue.
 
Agreed they need a rescan feature if Adobe plan to make improvements.
I'd also like to see them recognise faces in profile as well. I don't mind if the software can't suggest a name (maybe that's beyond the capabilities of the software) but it would be nice to at least see it as unidentified.
 
I also need to rescan. And there is a way. Highlight all the pictures in the filmstrip. Right-click and rotate CW. LR immediately rescans for faces. It won't find many, so right-click again and rotate CCW. This time all the faces will be rescanned and either marked as unknown or recognised. Bit of a pain, but it works. Hope this helps!
 
Welcome to the forum Stephen. Thanks for sharing that tip!
 
I started kinda going with the mis identification, and used faces for things that aren't faces, like pets. Figured what the heck? and it locates the keyword (which the names are essentially) in an area in the photo.

I'm kinda wondering if more keywords ought to work like faces, with location-in-image frame info. For some bird pictures it could be useful, although one needs an application or website that can make use of it. I wouldn't have to put info in the caption like "kingfisher on the right, pelican on the left." But I haven't committed to cramming all that into "Faces" because it could be mess to undo.

And BTW if you crop someone with an ID'd face out of the image in Lr and then export, the keyword for that face remains. Lr isn't smart enough to delete it. Could be embarassing....
 
I have found each dot release is a little better in terms of the facial recognition, both correctly finding faces and identifying the person.
 
I have found each dot release is a little better in terms of the facial recognition, both correctly finding faces and identifying the person.

Is that a new algorithm, or is it that Lr just re-searches upon update, and with access to your pre-existing picks, finds more faces more accurately? I've been wondering how much training affects face recognition, a la OCR software.
 
Short answer is I do not know definitively. I can say from my perception, each release has done a better job of correctly identifying what are actual faces on the initial scan. In addition once it finds a face, Lr does a better job of guessing which person it is. This is both for those that I have many images on, and new people I have never tagged before.
 
Status
Not open for further replies.
Back
Top