LIGHTROOM CC - Facial Recognition is really great for awhile, then gets corrupted

  • 2
  • Problem
  • Updated 4 years ago
  • (Edited)
It took about 36 hours to complete indexing faces on my 80,000 photos. Initially the recognition was REALLY excellent. Once I named about 10 people (my family members) it was going very well, and very accurately suggesting the right names ...

And then, it seemed to die... I got the dreaded "Lightroom is not responding.." and I had to restart it. When I started Lightroom CC again, it seemed fine. Just picked up where it was... but the Face recognition is now awful. It is suggesting the wrong names for EVERY face! It's as if the face information became corrupted.

Is there any way to rebuild the facial information based on the people I have already named? This is frustrating because what was going fairly smoothly is now very tedious and time consuming because the facial recognition is failing now.

Any thoughts or suggestions? There do seem to be some performance issues as it goes through the indexing , it seems to slow and slow and eventually become nonresponsive.

PC / Windows 7 / 32gb RAM / NVIDIA Geforce 560 GPU
Photo of Briand Sanderson

Briand Sanderson

  • 9 Posts
  • 0 Reply Likes
  • frustrated

Posted 4 years ago

  • 2
Photo of Simon Chen

Simon Chen, Principal Computer Scientist

  • 1490 Posts
  • 471 Reply Likes
Does your system has a spinning disk hard drive or a SSD?

The suggested names are not final. If you name, confirm more unnamed people, does correct suggestion start to show up?
Photo of Briand Sanderson

Briand Sanderson

  • 9 Posts
  • 0 Reply Likes
SSD - It has 45 GB free

When I first started indexing (80,000 photos so it took some time), the names were amazingly accurate. It was impressive.

Now, after completing the indexing of the faces, I am going through the suggested names. It's not getting any better. It does adjust as I add more confirmed people, but it is suggesting the wrong names for confirmed people and now its wrong most of the time. Yes, there are some times when it gets it right, but that's not as often as it was initially.
Photo of Simon Chen

Simon Chen, Principal Computer Scientist

  • 1490 Posts
  • 471 Reply Likes
Too much read/write during indexing could be problematic for SSD. I don't have hard evidence but I heard something like it before.

With more faces detected, name suggestion could appear weaker because it processes a subset of faces at a time. Wrong suggestion could also happen if you have some mis-labeled named faces.

Two suggestions:
1. Double click on a named face (or unnamed faces work as well) to go to the Single Person View, Lr would show all instances of the named faces and all similar looking faces at the bottom pane. Check if the named faces are labeled correctly, and label the similar looking faces for the same person there. It is much more efficient this way if you are only interested in labeling a few known persons.

2. Try labeling with a smaller set of folders/collections (instead of All Photographs) at a time, the face suggestion and name propagation will be more targeted and accurate that way.
(Edited)
Photo of Briand Sanderson

Briand Sanderson

  • 9 Posts
  • 0 Reply Likes
I think a lot of power users are using an SSD, so hopefully you're not saying that Lightroom CC users should not be using an SSD?

It seems I must completely not understand how Lightroom's facial recognition works. I assumed that the more faces you identify, the stronger the match will be for confirmed faces. So if I have confirmed 100's of faces, that the algorithm would get better and better at discerning what face belongs to what person. But it sounds like you're saying that is not how it works at all.

I tried your suggestions. Suggestion #1 didn't work well. Most of the time I had go hunting through 100's of "Similar" faces, and sometimes the set of Similar faces didn't even have the person I was identifying. I did ensure that all my named faces were labeled correctly. I did find a few that were mislabeled. It seems that choosing the check mark doesn't mean what I thought.. But even after cleaning up the mislabeled faces, I didn't really see any improvement to the algorithm.

Suggestion #2 allowed me to complete the task much faster. The algorithm really didn't improve, there was just a smaller set of faces that made it easier for me to drag/drop sets to the right confirmed face.

Perhaps my expectations are too high, but it would seem to me that if I have confirmed a collection of faces, with 100's of examples of a single person, that when I scan new photos the algorithm should be really good at detecting who is who. But that just doesn't seem to be the case here.
Photo of Simon Chen

Simon Chen, Principal Computer Scientist

  • 1490 Posts
  • 471 Reply Likes
I am not saying that SSD should not be used. It just that some SSD and driver could be buggy. I am not even sure your current issue is related to that.

The engine will learn as you confirm more faces from the same person, but you need to make sure you label them correctly.
Photo of HyunMo Koo

HyunMo Koo

  • 13 Posts
  • 1 Reply Like
It could be LR is working normally but only ran out of faces which are easily recognizable.

When I usually work with faces in LR CC, first 50-80% faces of the same person is easily recognized, but other 20-50% requires manual tagging. :(

And more you tag, more slow it gets. I think it's because LR has to compare similarities with more samples. It does get more accurate. But, accurately recognizable samples run out. (Because those are always tagged earlier in the process.)