Dante_Stella
Rex canum cattorumque
So the first mistake was setting LR6's face recognition module loose on 160,000 photos (many of which are scans of substantial size).
That said, the module seems to have some features, some good:
That said, the module seems to have some features, some good:
- It does not confuse ex-girlfriends with my wife.
- It recognizes my childhood dog.
- It does not confuse my childhood dog with my ex-girlfriends.
- It does not confuse my wife with my mother.
- It has a very high recall and excellent precision compared to iPhoto, Aperture, and Picasa.
- Unlike most machine learning systems I have used, this one never seems to reach a stable point. No matter how high the precision and recall, you still get tons of "unnamed" faces that actually have the right names on them. For a system like this to be effective, it needs a fault tolerance. The goal of this system should not be to reach 100.0% accuracy. It should be to get to a point where things are good enough to serve as a labor saving tagging device.
- It is dog slow to look at said "unnamed faces" display.
- For unnamed faces, you only have the choices of (a) accept or (b) delete the face box. You see that as "Dante Stella?" What the system needs is a "no" button that gets rid of the face until the analytics can come back with the next best match.
- Unlike other systems, you can see that each call is only resulting in the filing of that precise picture and any duplicates - NOT what you would expect, which is that as you update the data points, many things are going to update, which presumably includes things disappearing from the unnamed faces list.
