What is a good face detection workflow?

[daver99]

Michael Presley's comments here got me to wondering what a good workflow is to take advantage of the face detection features. I had it fully enabled for awhile, then turned off the background processing for performance reasons, but regardless of that side of things. When I did just "go for it" with tagging a few faces, but not many I then ended up with virtually every face coming in being tagged as one person and I would then have to correct them all. I just deleted all the face data with a plan to start again, but I suspect that to get the best, most automated, results I'd want to do something like this the following.. but I would love some feedback on what others have found to be successful uses of the new feature:
1. Turn off background scanning in manage mode
2. Use the view mode tools to name some number of instances of the key people that you want to system to know about (e.g. for me, my extended family and a few key friends)
-- Q: how many instances do you need to tag for the system to build good metrics for differentiating people?
3. Once you've ensure a good number of photos are all correctly scanned for all your key faces, THEN turn on background scanning and just leave it open all night..

.. or am I out to lunch on this?

A best practices guide would be awesome.

DR

[Tristan H]

Hi daver99,

This is a great workflow that should definitely help control incorrect names. I've posted about it before on the forums, but the more it gets mentioned the better.

We've found that even one correctly labeled face for each person will significantly help future recognition. More is always better, but 3-5 labeled faces per person should be plenty to help recognition get started. Maybe a few more in different lighting, side profiles, people wearing sunglasses, different ages, or people who look quite similar (e.g. close family members).

Thanks,
Tristan H.
ACD Systems

[daver99]

Hi again,
I finally got back to checking how my whole collection was tagged up after training the system with 10-20 or so images for each family member. Overall, the results were good in some ways, but have a big shortcoming.

The system seem to want to assign a known name to every face it sees - rather than leave anything as "unknown" so almost every person who had not been named before is now listed as my son Ben. I think then also once it has taken in the nearest match, the system expands the feature space for what it knows as "Ben" and then more unknowns will fall to be Ben, which in turn expands the feature set of BEN and now it seems BEN is most faces.

The results look like a feedback loop gone awry. I suggest the criteria for a match should be tighter and ideally there would be a big list of "unknowns" you could look to and use to train on new faces.

dR