LR CC Facial Recognition Sucks

[DAYMX5]

Started using the new facial recognition and it is a real crap shoot. Saying it gets it right 50% of the time would be generous. FaceBook is great compared to this.

I've spent more time correcting it then I would have doing it manually.

 

[BenjaminSpencer.com.au]

I'm finding it really good!

Facebook only starts suggesting names after it has a lot of tags of a specific person and it learns what people look like.
Lightroom has to do the same before it becomes accurate, it just starts guessing earlier - try tagging 5-10 photos of a specific person, then go to that person in the named persons section and look at the similar faces

You can then bulk select and confirm the top ones that are correct, and as you add more and more it gets more and more accurate.


At the moment Lighroom has finished searching 2/3rds of my ilbrary of 75,000 images
Found 14,197 unnamed faces, 4661 named faces

 

[DAYMX5]

Tried it quite a few ties and it never got any better. It my learn but apparently it forgets easily!!!!!

 

[gregDT]

I found it improved over time quite a lot. Once it had scanned a folder and had logged names to faces it began to pick them up with increased accuracy as new folders were searched. I was also pretty impressed that even though a few people featured had aged from about two years old to over eight it could still often recognise them. Adobe's face recognition isn't as good as whatever was used in Picasa though. That software was frankly scary!

 

[Piggy]

I agree with the OP. It pretty sucks. My purpose in using facial recognition is to tag photos that contain images of people I'm interested in for the purpose of finding images with these people. i.e. Family and friends. Show me the photos with Tom and Jerry from last month's family vacation, etc.

I have 8,500 photos. I ran recognition on all of them. 6,000 faces were identified. I've invested several hours of review and have about 4,000 faces that still need to be reviewed. LR entirely misses identifying a face is present in many photos where it appears very obvious to me - nice contrasty background yet it finds them in framed photos that are small and appear in the background of some interior shots. This makes no sense to me. It also identifies a fare amount of sections of photos that are not even faces, like walls. Another problem I've had is it identifies a stack of different people as the same person but presents them as different people when the stack is expanded. e.g. It presents a stack of 3 Johns. I expand the stack and the individual photos are John, Ted and Jill - and they really are 2 or 3 different people. WTF? If LR thinks they are different people (which they were, even though not necessarily guessed correctly) then why is it stacking them as the same person?!

I realize this is the first release of LR facial recognition but this is not a new technology by any means. I expected better. I think I need to stop before wasting any more time with this feature.

This feature should have been released as a public beta!

 

[tspear]

I disagree. I have about 18K images, Lr found about 20K faces. Since I mostly take pictures of family, friends... and rarely just landscapes this makes sense.
I have gone through about 3K images so far; I would say Lr is about 80% accurate on the guessing of the name so far. Where it makes a mistake, it often is sisters or father/son... And they look alike.
For me, this is better then Apple Aperature, but not as good as Google Picasa.

Tim

 

[clee01l]

The database entries for facial recognition have been in the LR catalog since version 4. This tells me that Adobe has been working on this function for some time before releasing it to the user base. However, this is a first release and as such, my expectations are lower than those for a mature feature. I find facial recognition useable, and will improve with user feedback. My expectations are that facial recognition updates from Adobe will sharpen the results that you are seeing presently.

 

[Julius Mandelblat]

IMO, this capability is not ready for use at all. It completely ignores people that are viewed from the side (profile), some other angles of view or wearing sunglasses. It shouldn't be judged by only tagging the people it finds, but also by how much it misses to find faces at all. Its accuracy is also completely inadequate and it continues to do very rough mistakes (mixing a guy with beard with a young girl) after getting thousands of images to learn on. In short - in its current state it is useless. It is heavily behind the industry state of the art. It can't be compared with long existing face recognition capabilities of Google or Facebook. It is a shame that so professional software company releases such poorly designed feature.

 

[Kirby Krieger]

It's a ... bonus feature. Most people judge it against their expectations, which may provide some emotional pleasure but is not helpful for work.

It does some things well. It does others just OK.

It has almost certainly been marketed as more useful than it is.

Do the things it does well save you time? Use them.

Overall does it cost you more time than it saves? Turn it off and ignore it until Adobe presents an overhauled version.

If you use it, come to grips with the difference between "Photo contains named face-region" and "Photo depicts person".

Personally, I like Lightroom's People View a lot. I think the People/Person UI is well done, as is the implementation within and integration with Lightroom.

In my use, I almost never add a face region (or whatever Lightroom calls it). I don't use People view to tag people, I use it to tag faces.

I'm not sure what the OP means by

I've spent more time correcting it then I would have doing it manually.

"Correcting" implies changing something that was set. Lightroom suggests a Face-Name (again, not sure of Lightroom's nomenclature) for a region it thinks may be a face. If the suggestion is correct, I click the check mark. If the suggestion is wrong, I drag the thumbnail to the correct already-saved Face-Name represented by a thumbnail at the top of the work area (the "Named People" section). If the region is not a face (or I don't want Lightroom to name and remember this region as a Named Face), I type (press and release) the {delete} key on the keyboard.

In practice this is quick. When I have scores of suggested Face-regions in the work area, I usually first ⌘-click all the regions I want to dismiss (creating a selection of multiple suggested Face-regions), and type (press and release) {delete}. Then I identify — either by checking a correct suggestion or filling in my own — one of each person in the work area whose face-regions I want to tag. After that it is simple work to either check correct suggestions or drag incorrect ones (keep the suggested Face-regions sorted by "Suggested Name" — this is set on the Toolbar).

Remember well: "Named Face" _is a subset of_ "is person in Photo" — specifically, the Photos that have been given a Face-region assigned to a People keyword.

Imho, all the tech companies do users a disservice when they suggest an equivalence (and not a subset). _If_ your work involves _faces_, face-recognition can be handy and a time-saver. If your work involves tagging, grouping, and sorting by persons, I suggest simply using regular keywords.

My work involves faces. I am well-served by the presentation of many "head-shots" in People View — it lets me quickly scan dozens, scores, hundreds of Photos in order to select by facial expression.

NB: The Lightroom Queen's section on this is exemplary (The Missing FAQ, p. 151ff).

 

[tspear]

Kirby,

That is pretty much what I do. Except I do not "drag and drop" to a named individual. I just type the name in the box.

 

[Julius Mandelblat]

It's a ... bonus feature. Most people judge it against their expectations, which may provide some emotional pleasure but is not helpful for work.

It does some things well. It does others just OK.

It has almost certainly been marketed as more useful than it is.

Do the things it does well save you time? Use them.

Overall does it cost you more time than it saves? Turn it off and ignore it until Adobe presents an overhauled version.

If you use it, come to grips with the difference between "Photo contains named face-region" and "Photo depicts person".

Personally, I like Lightroom's People View a lot. I think the People/Person UI is well done, as is the implementation within and integration with Lightroom.

In my use, I almost never add a face region (or whatever Lightroom calls it). I don't use People view to tag people, I use it to tag faces.

View attachment 9144

I'm not sure what the OP means by


"Correcting" implies changing something that was set. Lightroom suggests a Face-Name (again, not sure of Lightroom's nomenclature) for a region it thinks may be a face. If the suggestion is correct, I click the check mark. If the suggestion is wrong, I drag the thumbnail to the correct already-saved Face-Name represented by a thumbnail at the top of the work area (the "Named People" section). If the region is not a face (or I don't want Lightroom to name and remember this region as a Named Face), I type (press and release) the {delete} key on the keyboard.

In practice this is quick. When I have scores of suggested Face-regions in the work area, I usually first ⌘-click all the regions I want to dismiss (creating a selection of multiple suggested Face-regions), and type (press and release) {delete}. Then I identify — either by checking a correct suggestion or filling in my own — one of each person in the work area whose face-regions I want to tag. After that it is simple work to either check correct suggestions or drag incorrect ones (keep the suggested Face-regions sorted by "Suggested Name" — this is set on the Toolbar).

Remember well: "Named Face" _is a subset of_ "is person in Photo" — specifically, the Photos that have been given a Face-region assigned to a People keyword.

Imho, all the tech companies do users a disservice when they suggest an equivalence (and not a subset). _If_ your work involves _faces_, face-recognition can be handy and a time-saver. If your work involves tagging, grouping, and sorting by persons, I suggest simply using regular keywords.

My work involves faces. I am well-served by the presentation of many "head-shots" in People View — it lets me quickly scan dozens, scores, hundreds of Photos in order to select by facial expression.

NB: The Lightroom Queen's section on this is exemplary (The Missing FAQ, p. 151ff).

I do not understand what is meant by "bonus feature", at least Adobe doesn't grade it like that. Even in your example there is no a single face from side view. Being familiar with the state of the art of such technologies I can only give it "less than satisfactory" grade.

 

[clee01l]

I think you need to look at facial recognition as a feature implemented by Adobe to compete with the social media photo apps to please the "selphie" user. To that end, it does a pretty good job. If you are a serious photographer, using a serious camera, and don't concentrate on (self) portraits, then you probably don't need or want to use facial recognition. If you are a professional photographer, taking session portraits, then facial recognition is redundant. If you are a professional photographer taking wedding or sporting event photos, then LR's 1st generation facial recognition may be adequate although most individuals may only appear once in only one frame. Before condemning LR's 1st generation facial recognition feature, look at your image inventory and ask your self how many faces did the facial recognition feature miss? Not how many "faces" that it found that weren't faces at all. The facial feature is an iterative learning engine. The more faces that you correctly identify, the better LR becomes at predicting correctly. One of the thing that I discovered when using facial recognition is how close a resemblance family members had with each other. When I identified my own face to LR, LR quickly suggested faces of my son, my aunt and my male & female cousins, even though I saw no resemblance at all. Once I trained LR to recognize my aunt and my cousins as individuals, LR made fewer wrong family guesses.

Taking the time to train (assign a name) LR's facial recognition is a serious effort requiring significant up front time. You get out of LR's facial recognition only the benefits corresponding to the effort you put into training the function. I stopped using the function for that reason. Less than 5% of my photos have people in them that I know, So why bother with the processor overhead when so few photos that I take have people in them and even fewer have people in them that I know or care to know.

 

[Kirby Krieger]

I do not understand what is meant by "bonus feature"

My wording was clumsy. I should have said "I think it best to think of Lightroom's facial recognition as a bonus feature". I'm trying to get users to focus on whether it is useful to them, and if so, how to best segment it into their workflow.

I can only give it "less than satisfactory" grade

I agree with your grading, but the grade and grading is at best obliquely helpful to Lightroom users. The question isn't, imho, "Does it suck?"; the questions we should answer are "What does it do? How can this be used? Is it worth using it?" Cletus, above, gives an excellent summary introduction to Lightroom's facial recognition and it's utility.

 

[tspear]

I find it useful, and it is in my workflow. Not sure if because of training or they have improved over the releases, but it is getting better.

 

[Linwood Ferguson]

Did anyone find a way to say "I just want to start completely over and wipe out everything"?

If it might be better I don't want the ones I've done. In particular, what concerns me are the images it did NOT find faces in, and marked as processed.

Or do I need to figure out some SQL?

 

[Kirby Krieger]

In particular, what concerns me are the images it did NOT find faces in, and marked as processed

You might find it interesting that Lightroom re-processes for face detection any Photo that is rotated.

Which _seems_ to indicate that the "processed for face detection" record is a toggle.

Fwiw, I do four passes through all Photos in which I wish to have Lightroom detect faces: upright, rotated 90°, rotated 180°, and rotated 270°. I apply a keyword to each Photo for each processing pass.

"I just want to start completely over and wipe out everything"

What I mention above might suggest, if you want to "start over", that you delete all "Person" keywords* and then rotate all your images 90° in one direction, and then 90° in the opposite direction. If you check the status of your catalog by clicking the name in the Identity Plate you'll see that Lightroom is reprocessing all your Photos. (You can see additional confirmation that Lightroom is processing for face recognition by going to People* view.)

____________________________

*I want to state, strongly, that "People" is a bad name (along with "Person") — either ignorant or malignant — for the results of face detection. Face detection does not detect people — it detects (some) faces. A "person" detector would be — for me at least — a very useful tool, the results of which would be quite different from the the results of a face detector. Conflating the two — nay, substituting the grand "People" for the plainly expository "Faces" — confounds users, raises expectations, and implicitly makes a claim that cannot be met.

Why not just call the results of face detection "faces"? Is this the diktat of the same Minister of Nomenclature who thought it good to call a database a Catalog?

 

[Linwood Ferguson]

What I mention above might suggest, if you want to "start over", that you delete all "Person" keywords* and then rotate all your images 90° in one direction, and then 90° in the opposite direction.

Thanks, good to know, though I'd need to do some testing to make sure the net-zero rotation does not cause them to be re-uploaded in a publishing plugin. Uploading 40,000 image or so to Smugmug would take a while. :eek2:

I'll dig into the database when I get some time. I bet this is an easy thing to clear out with a bit of SQL.

(Sound of Jeopardy Them Song playing)

OK, took a brief look...

There's a separate table (Adobe_libraryImageFaceProcessHistory) that contains status information, interestingly including a "lastFaceDetector" and "lastFaceRecognizer" as though there are (or may be later) multiple types or versions. I do not see a base table for those definitions (mine are "2" and "3" respectively). Also interesting with those is a lastImageIndexer which is always null, but might indicate something planned (or just not implemented). A few of my lastFaceDetector are -2 and those have no lastFaceRecognizer, so it may indicate some kind of work-ahead queue.

There are then tables for agLibraryFace, AgLibraryFaceCluster and AgLibraryFaceData. FaceCluster contains a keyFace, but all of mine are empty. FaceData contains a bunch of binary data, my guess is that's the recognition data in some form, and/or what it shows as "the" face image. The Face table has references to image and lots of positional information.

This would take some experimentation (e.g. snapshot, change something, compare) but at first glance it looks like you could just wipe all this out en-mass, and leave the keywords, or perhaps even leave the accumulated face definitions (i.e. how it recognizes) while reprocessing images, without forcing a republish (except indirectly of course if you add face keywords). Maybe. Big difference in reading and speculating and actually changing the catalog, the latter requires a lot more care.

I'm really surprised I haven't seen more plugins (which would be a safer path) to do this -- maybe Adobe has not exposed any of the facial recognition data in the plugin interface -- anyone a plugin author around here and know?

 

[Julius Mandelblat]

I think you need to look at facial recognition as a feature implemented by Adobe to compete with the social media photo apps to please the "selphie" user. To that end, it does a pretty good job. If you are a serious photographer, using a serious camera, and don't concentrate on (self) portraits, then you probably don't need or want to use facial recognition. If you are a professional photographer, taking session portraits, then facial recognition is redundant. If you are a professional photographer taking wedding or sporting event photos, then LR's 1st generation facial recognition may be adequate although most individuals may only appear once in only one frame. Before condemning LR's 1st generation facial recognition feature, look at your image inventory and ask your self how many faces did the facial recognition feature miss? Not how many "faces" that it found that weren't faces at all. The facial feature is an iterative learning engine. The more faces that you correctly identify, the better LR becomes at predicting correctly. One of the thing that I discovered when using facial recognition is how close a resemblance family members had with each other. When I identified my own face to LR, LR quickly suggested faces of my son, my aunt and my male & female cousins, even though I saw no resemblance at all. Once I trained LR to recognize my aunt and my cousins as individuals, LR made fewer wrong family guesses.

Taking the time to train (assign a name) LR's facial recognition is a serious effort requiring significant up front time. You get out of LR's facial recognition only the benefits corresponding to the effort you put into training the function. I stopped using the function for that reason. Less than 5% of my photos have people in them that I know, So why bother with the processor overhead when so few photos that I take have people in them and even fewer have people in them that I know or care to know.

The answer on your question is simple. This feature misses ALL profile views of the faces and thus useless.

 

[clee01l]

The answer on your question is simple. This feature misses ALL profile views of the faces and thus useless.

As stated elsewhere, This is Adobe's first venture into Facial recognition and this was mainly done to compete with the social media apps and not really intended as a useful tool for serious photographers. It has been my experience with the tool that the function does a pretty good job of identifying faces as long as there are two eyes and a mouth in the general normal anthropomorphic position. A 3/4 view and full faces usually fit this criteria. Side views (profiles) or back of the head views do not and as you indicate are not detected. So, does Adobe need to add this in version 2? A lot depend upon how much use is actually being made of the existing facial recognition and whether there is a significant number of peopl using the feature and are they clamoring for side view recognition or recognition of the back of the head? At the top of this page is a link to Adobe to report bugs and to put in a feature request for features that you feel are missing. If there existing bug reports that cover your issue, you can add your support to these. Adobe is more likely too give a feature request attention of several people have the same problem and support each others request

 

[Kirby Krieger]

This feature misses ALL profile views of the faces and thus useless.

I'm not following your logic ... unless you are presuming that your needs are universal?

 

[mrgooch]

Started using the new facial recognition and it is a real crap shoot. Saying it gets it right 50% of the time would be generous. FaceBook is great compared to this.

I've spent more time correcting it then I would have doing it manually.

 

[mrgooch]

Is there a way of deleted the app?

 

[clee01l]

Is there a way of deleted the app?

Facial Recognition is a feature of the LR app. You can turn the feature off if you don't need it or don't want to spend the time to train it.

 

[SnrSachse]

Professional portrait photographers may have no use for the feature but if they are the only ones this software was designed for then Adobe needs to stop advertising as a tool for the rest of us because we are spending our money on it just like everyone else. The Amazon Prime free photo cloud face recognition is an order of magnitude better and that was on their first pass. The free Picaso software did a much better job and that was years ago. If LR want to sell themselves as a both an editing tool and an organizational tool than they need to get much better at the origination side and get away from the belief that they can just cater to a few portrait photographers who really only need to keep the sessions separate and recognize the serious amateurs who use it and want a single tool that can handle the family photos and their more serious stuff. I am not a professional photographer but I am an avid photographer and I have tens of thousands of photos spanning a couple decades and I need a tool that simplifies organization and can do quality editing. LR is amazing at editing but the organization tools are less than adequate, face recognition is just the worst of a bad tool. Editing needs to be complex but organization should intuitive and easy to use. I spent hours training the software and it is still misses my daughter's face half the time even though it has thousands of samples going back eleven years and doesn't just confuse her with relatives. I have seen photos that were taken back to back with clear faces looking straight at the camera and it will assign these almost identical images to two different names. Worse yet even after hours of assigning faces the tool simply isn't that useful or intuitive and really does not help that much with sorting photos.

They need to fix it or junk it because so far all they have accomplished is wasted hours of my time on a tool that is one of the worst I have seen in this field. LR and PS are difficult turns to learn but usually well worth the time and effort it takes once you have learned a new skill, that was not the case with this feature and it is the wasted time that bothers me the most. The tools is worse than useless.

 

[LRList001]

Professional portrait photographers may have no use for the feature but if they are the only ones this software was designed for then Adobe needs to stop advertising as a tool for the rest of us because we are spending our money on it just like everyone else. The Amazon Prime free photo cloud face recognition is an order of magnitude better and that was on their first pass. The free Picaso software did a much better job and that was years ago. If LR want to sell themselves as a both an editing tool and an organizational tool than they need to get much better at the origination side and get away from the belief that they can just cater to a few portrait photographers who really only need to keep the sessions separate and recognize the serious amateurs who use it and want a single tool that can handle the family photos and their more serious stuff. I am not a professional photographer but I am an avid photographer and I have tens of thousands of photos spanning a couple decades and I need a tool that simplifies organization and can do quality editing. LR is amazing at editing but the organization tools are less than adequate, face recognition is just the worst of a bad tool. Editing needs to be complex but organization should intuitive and easy to use. I spent hours training the software and it is still misses my daughter's face half the time even though it has thousands of samples going back eleven years and doesn't just confuse her with relatives. I have seen photos that were taken back to back with clear faces looking straight at the camera and it will assign these almost identical images to two different names. Worse yet even after hours of assigning faces the tool simply isn't that useful or intuitive and really does not help that much with sorting photos.

They need to fix it or junk it because so far all they have accomplished is wasted hours of my time on a tool that is one of the worst I have seen in this field. LR and PS are difficult turns to learn but usually well worth the time and effort it takes once you have learned a new skill, that was not the case with this feature and it is the wasted time that bothers me the most. The tools is worse than useless.

My experience of Adobe's face recognition has also been disappointing. When I first used it, it was astonishingly good, in fairness it was terrific at spotting matches of the same person at different ages and all sorts of head angles. Then bit by bit as it matched more and more images of that person, it gradually fell apart, to the point where it started to give absurd matches, so I stopped using it. For me, Adobe's face recognition has ended up being "worse than useless" too.