Face Reindexing... v7.3. Still needs work.

[philc30713433]

Great, it found more faces...  It didn't stack them.  I have model shots that are literally 15 images taken moments apart with the model in the same position, and I have to SHIFT / CTRL click them to lump them under the person.  If I want to find the missing folks in thousands of images, I literally have to retag EVERYONE??  I am really disappointed. 

Why does the stacking not work?  I actually took a subset of my photos, through them in a new catalog and reran the facial recognition on it.  Some of the stacks were 100 images deep, however in my main catalog where I add images week-by-week, I have hundreds of one-off images.  It's like it isn't updating the learning algorithm and reapplying it to the images in the catalog.  Or it gets so overwhelmed that it's trying to do something O(N^3) or god forbid O(N^N-1) time!

Simon, did y'all only retool the algorithm and let us re-run it?  It still seems as problematic as before that it is not grouping images.  In a catalog with 3,000 images with no existing tags, I re-ran the image search.  Instead of 3,800 faces it found 5,000.  However the biggest stack is only 5 photos deep!!  I literally will have to click approximately 3,000 times to identify all these faces and that's if I don't have to type in a bunch of new people which I do.

Also, can we talk about the fact that the lookup doesn't like the " " (space) character so "FirstName LastName" for a name like David pulls up every David in the catalog, and if I start typing a last name with an S I then get every "Sam", "Samantha", "Stephen" etc instead of "David S" continuing to winnow down.

This is somewhat (very) disappointing for a feature that we've literally been waiting years for an update on.  Also, no word on pausing it or why my processor is constantly at least 25% CPU for Lightroom Classic.  It sounded like this was the image/facial recognition churning, but I'm still getting stacks 5 images deep so what's the processor doing with all the input?  Also, why can't we stack like faces in the Unnamed section?  Like if I have 15 images sitting next to one another clearly of the same person, Yes I can SHIFT click or CTRL click and name them all simultaneously, but if I just use a better facial recognition service these are all easily grouped.

Anyone seeing anything else?  Would love to get a comprehensive list of items they can use as an improvement list.

[philc30713433]

I let this run overnight and still the biggest stack size is only 5 photos.  Something is seriously wrong that it can't identify model shots (seconds apart from one another) as the same person in about 12 hours of runtime.

[99jon]

As you discovered, the new engine finds more faces but starting from scratch it will involve manual work. I guess that is why Adobe provided the overrides for those who have already run the indexing. A checkbox has been added to allow you to retain all confirmed faces (confirmed named faces or manually drawn faces) when opting for “Find Faces Again”. I know this does not help you but personally I would not use this for model shots. It is far easier to keyword import batches.

[philc30713433]

I primarily use it for a weekly bar event where I see the same folks week after week over time.  I have 4 years worth of photos, approximately 13,500 images.  They are group shots primarily.  So this update is basically useless for me.  If an African American was missed by the first algorithm, I have to choose whether to lose all tags and cross my fingers that he is picked up in the second run, and then retag everyone.  This would be potentially worth it if the stacking were working properly.  Given that it is not, I don't want to retag 19,000 faces that I've already confirmed.