[Digikam-devel] Re: face detection process...

[Marcel Wiesweg]

> - 133 unknown items found.
> - 5 items are not faces. All others are fine.

Judging this result of course depends on the "true positive" number: How many 
pictures truly have a face?
The important part here, for use in digikam, is to run real-world tests: A 
collection where only 10-20% have faces at all.
The initial implementation gave me lots of false positives. And this gave a 
bad first impression: almost all found entries were not faces. Users will see 
and judge that it doesnt work at all. We must have most entries correct there.
Why the difference? The Positive Predictive Value of a test depends on the 
prevalence!

The current implementation is a two step approach:
1) With a single cascade, faces are scanned
2) With a larger number of cascades, faces are verified.

Step 1 determines the upper limit of sensitivity. Step 2 improves specificity 
and decreases sensitivity (but in my experience, not much)


> 
> Bugs :
> - when scan is complete, progress bar continue to move...

Does it display a percentage number?

> - There is an album displayed as Unknown where there is no item. Why ? Look
> screenshot for details. in fact, it's an icon view problem at scrolling.
> Moving contents with mouse wheel change icon view.
> 
> http://www.flickr.com/photos/digikam/5322873069
> http://www.flickr.com/photos/digikam/5322873101
> http://www.flickr.com/photos/digikam/5323478176

This is a model-view problem. I have seen it in a few cases, usually a problem 
with sorting. No idea atm, did never seen that with faces.

Marcel