Lightroom Classic desperately needs the Sensei version of face recognition

  • 4
  • Idea
  • Updated 2 months ago
Is there any chance we are ever going to get face detect in Lightroom Classic that comes anywhere close to working like it does in Lightroom CC? I was excited when face detection was first introduced because I could really benefit from it, but I almost immediately gave up on it because while the actual detection of faces is great, it is terrible at finding similar faces. Imagine my surprise today when I starting playing around with the same feature in the CC version at it is amazing. With no training on my part it had correctly identified dozens of individual people. It was 100% correct when it though photos were of the same person across very dissimilar photos. All I had to do was give each person a name.

Now lets compare that to a couple samples from Lightroom Classic. All of these photos are its guesses for a match to a 7 year old white girl. She already has 1500 photos confirmed as her so I can't see this getting any better over time.

 



I mean, seriously, how is this even trying? A see at least 4 different ethnicities, a 60 year age range, a dog, and a chain link fence. For the record none of these are the girl in question.

I am never going to use Lightroom CC as my primary photo editor. I can't afford the $1300/year it would take to store my entire photo collection in the cloud at the current costs and as a professional photographer I am not going to trust the primary copy of my photos to a server that I don't have control of. That being said, Adobe is putting some amazing technology in Lightroom CC and I would love to be able to take advantage of it. As it is, I can't even cycle photos through the cloud since non of the keyword syncing goes back to my Lightroom Classic catalog.

Photo of Jonathan Goforth

Jonathan Goforth

  • 11 Posts
  • 3 Reply Likes
  • frustrated

Posted 2 months ago

  • 4
Photo of Roelof Moorlag

Roelof Moorlag

  • 192 Posts
  • 57 Reply Likes
I’m completely with you!
Photo of Johan Elzenga

Johan Elzenga, Champion

  • 2167 Posts
  • 897 Reply Likes
It’s unlikely that this will happen. Sensei runs in the cloud, not locally, so without having your images in the cloud, Lightroom Classic can’t use it. The local algorithms of Lightroom Classic could perhaps be improved using Sensei, but I wouldn’t hold my breath.
(Edited)
Photo of Roelof Moorlag

Roelof Moorlag

  • 192 Posts
  • 57 Reply Likes
Excire brought AI to the Lightroom desktop, why should Adobe could not? 
A plugin like Cloud tagger brings images to the cloud, doing the AI there and brings back the keywords to the desktop, why should Adobe could not?
I think, it's a delibarate choice, not technical limitations.
(Edited)
Photo of Johan Elzenga

Johan Elzenga, Champion

  • 2167 Posts
  • 896 Reply Likes
I didn't say it could not possibly be done, I explained why I don't think it will be done.
Photo of Yves Crausaz

Yves Crausaz

  • 452 Posts
  • 62 Reply Likes
Une idée, sur LR Classic, créer une collection, la synchroniser avec LR Web, les aperçus dynamiques ainsi créés, ne bénéficient-ils pas sur LR de la reconnaissance Sansaï, pourquoi, les programmeurs d'Adobe ne pourraient-ils trouver une solution pour simplement compléter un champs des métadonnées donc le contenu serait simplement ramené vers LR Classic, ils y arrivent bien pour les modifications de développement, pour les champs "Titre" et "Légende" ! C'est certainement de la mauvaise volonté ou simplement ne pas écouter ses propres clients !

An idea on LR Classic, create a collection, synchronize it with LR Web, dynamic previews so created, do not on LR recognition Sansai, why, Adobe programmers could not find a solution for just complete a metadata fields so the content would be simply reduced to LR Classic, they do well for changes in development for fields 'Title' and 'Legend'! It's certainly ill will or simply does not listen to its own customers!




Photo of KWOK KAN HO

KWOK KAN HO

  • 2 Posts
  • 0 Reply Likes
I feel the same as Jonathon.
Photo of John R. Ellis

John R. Ellis, Champion

  • 4531 Posts
  • 1213 Reply Likes
"I was excited when face detection was first introduced because I could really benefit from it, but I almost immediately gave up on it...Imagine my surprise today when I starting playing around with the same feature in the CC version at it is amazing. With no training on my part it had correctly identified dozens of individual people. ... Now lets compare that to a couple samples from Lightroom Classic. All of these photos are its guesses for a match to a 7 year old white girl. She already has 1500 photos confirmed"

I don't know how the face recognition in LR Classic compares to LR CC, but this comparison isn't going to shed much light on it, unfortunately:

Since face recognition was first introduced, the underlying recognition engine was replaced in LR 7.3.  It's never been completely clear how LR handled the upgrade from the old engine to the new, and in particular, how well it incorporated the data from previously recognized faces into the indexing of new faces.  I think it's very possible, perhaps likely, that if you made a copy of your catalog and discarded all existing confirmed faces and reindexed the faces, recognition across your photos would be significantly different.

You're comparing the initial recognition in CC of the first dozens of instances of a person with your Classic catalog in which you've already confirmed 1500 photos, which isn't apples-to-apples. The machine-learning algorithms for face recognition produce an estimate of the probability that faces match. The recognizers in LR Classic and CC order matches presented to the user by that probability, the most likely matches first. So in your CC test, in which you've done little tagging, you're seeing the most likely matches for the girl being presented first, while in your Classic catalog, you've already tagged 1500 instances of the girl, so you're necessarily seeing possible matches that have much lower probabilities (that is, they're much less likely to be actual matches).

A more accurate comparison of CC with Classic would use the same exact corpus of photos in a fresh library / catalog.

Photo of John R. Ellis

John R. Ellis, Champion

  • 4531 Posts
  • 1213 Reply Likes
I've read nothing to indicate that the face-recognition engine in LR CC is different than the one in LR Classic. I also haven't read anything that indicates that a cloud-based implementation would be inherently better or easier to implement than a desktop-based one. 

There are three components to recognition algorithms: Identifying the bounding rectangles around the faces, extracting "features" from an individual face, and comparing those extracted features with the features of already tagged sets of faces in the catalog / library.  None of these inherently get any benefit from running in the cloud.

If anyone has links providing authoritative information about the face recognition in CC and Classic, I'd love to learn more -- it's entirely possible I'm wrong. While I'm not an expert in face recognition, I have over 15 years of experience applying machine learning in various consumer-product domains, in startups and huge companies.
Photo of Roelof Moorlag

Roelof Moorlag

  • 192 Posts
  • 57 Reply Likes
I have no real proof but in the tests i did the results in the cloud were substantial better than on my desktop. However it is always good to do a representative comparison.
Photo of Yves Crausaz

Yves Crausaz

  • 452 Posts
  • 62 Reply Likes
I'm thinking ... How can Sansa  find people by name, first name or any other personal criteria ????    will be magic!

Photo of Roelof Moorlag

Roelof Moorlag

  • 192 Posts
  • 57 Reply Likes
I'm thinking ... How can Sansa  find people by name, first name or any other personal criteria ????    will be magic!

It wil only find similar faces, you have to name them yourself