#Facial recognition feedback

19 messages · Page 1 of 1 (latest)

timber python
#

Would anyone like to give feedback on where facial recognition makes mistakes, if any? Try to think back on what corrections you had to make.

  • How often did it duplicate any people, if at all? Is there a pattern for which people this happened to, such as age, race, gender, face count, etc.?
  • How often did it merge more than one person together, if at all? Were they 2 people, or more than 2? If so, is there a pattern for which people this happened to, such as age, race, gender, face count, relation, etc.?
  • How often did it create two different people correctly, but mixed them up when assigning their faces? If so, is there a pattern for which people this happened to, such as age, race, gender, face count, relation, etc.?
  • If you made any corrections, was the facial recognition then accurate for future images?
  • Did you change any recognition settings from default? If so, what changes did you make and what effect did you notice from changing them, if any?
  • Approximately when did you import the library or last reset facial recognition (if ever)?

Context: I'm testing different facial recognition approaches and want to make it the best it can be.

If everything worked perfectly from the start, that's also useful feedback!

glossy cypress
#

In my experience, with default settings, it mostly tried to merge very young related children. After setting the distance a bit more strict it stopped doing that but instead created a lot more people from them. After merging them all it mostly stopped doing that apart from vague pictures.

timber python
long wind
#

I do not remember if I have modified it but here are my values and I do have some duplicates faces based on age sometimes

image.png
drowsy folio
#

#1452317522723344458 message @timber python ive written it in here to avoid cluttering? (i can copy paste it in here if you want to keep it in one place) (i can add more detail if you want or didnt get what i meant?)

glossy cypress
worldly junco
#

For me I notice multiple people for the same person on my initial import from when the person was an infant to an adult. This wasn’t like one for them as a child and one as an adult but I actually noticed multiple profiles could have young and older photos, with at least one being more focused on the younger years.

I also have the issue where i get a lot of different people created for babies/infants too, where each of these do seem to focus on one person, but includes quite a few pictures that aren’t them as babies. Usually it’s same gender, but occasionally mixes them too, especially around siblings/cousins.

Also large wearables to a face created a different person such as ski goggles and face paint.

These are all on the default settings, I’ve imported the library just last week and haven’t re-assigned anything into 3.2 comes out in case I’ll need to rerun it anyway.

river zinc
#

Adding onto the ski goggles, I had a person with about 20 normal photos and another person with 5 photos with ski goggles. The same person, so I merged them. Now, everytime the facial detection detects garbage as a face, it's merged into that person. And I this is something I notice occasionally, that things that are not a face are detected as such and merged with specific people.

worldly junco
river zinc
worldly junco
#

Huh weird

timber python
river zinc
timber python
river zinc
timber python
river zinc
#

Must have lowered it because it wasn't detecting enough faces though

timber python
#

ha well it's still not great that it blobs unrelated non-faces together. they would ideally sit alone and be hidden by the min face count threshold

honest berry
#

Eh, found this thread after posting my query in main channel

#immich message