The new Windows Live Photo Gallery beta greatly improves on people tagging, adding face recognition to face detection. My first look at this features finds some good news and also some of the other kind. Mind you, this is a beta version, so it is not the finished, polished product. Unfortunately, there is still a large amount of tweaking, adjusting, and correcting to be done in people tagging.
But first the good news. The older version of Photo Gallery was quite good at detecting faces that were face-on and level. The new beta version has greatly improved on this and now detects even most full side-on (profile) faces. Here are some examples:
Even faces in difficult lighting are detected. Here are couple of statues. One in harsh top light, the other in deep shade.
This is impressive progress. But now for the not-so-good news. At present the detection algorithm makes way too many mistakes. I provided Photo Gallery with about 7000 photos. It found 758 “faces”, sadly more than a third of these were “false-positives”, not faces at all. It also missed a lot of real faces. Here are some of the falsely identified faces:
The first two are indeed “faces”, but do not come under the category of “people”. The next three are a basket sculpture, tree branches reflecting in windows, and patterns in a leaf. Look closely and you can see the “faces”. Some other “misses” are harder to understand:
Can’t argue that the first is a face, and surely everyone recognized the one on the right!
So face detection has a ways to go. By way of perspective, the current version of Picasa detected 635 faces. It did not find all of them and did not do as well on profiles, however, it came up with just three false-positives. None of these three were found by Photo Gallery.
There were also a large number of faces that Photo Gallery did not detect. In this photo of the dancers there are twelve recognizable people (the reproduction here is too small to show them well).
Surprisingly, the face of the dancer in the foreground is not in the “Who is this?” list. Here is a screen shot from the “No tagged or detected people” listing.
The photos shown here all show people, some at difficult angles, but some in close up. Makes you wonder what the detection algorithm “didn’t see” in these.
In people tagging I discovered an overly eager process. I was tagging some old photos, one included a picture of a baby, and I created the new people tag “Jeffrey”. A bit later I noticed that the label now had a last name added (and an e-mail address assigned). In another photo I labeled “Dan” only to find that the program now insisted it was “danny”. The list of people tag options now contains my full contact list. There does not seem to be a way to remove my bank, car repair shop, and all the other businesses that I correspond with. As I said, a bit overly eager. Hope those horses get reigned in before long.
I have not (yet) tagged enough people to help Photo Gallery do well on face recognition. The instructions for batch tagging / face recognition are not yet there and the process is not that obvious. The program readily finds photos with faces when the people are in very similar poses.
It is not very often that we are afforded a look into the making of a powerful software product. Here we can appreciate the difficult task that the folks at Microsoft have set for themselves. It is not “the computer” that can tell the difference between a pattern in a basket and a human face, that has to be done with programmer-created algorithms that analyze patterns of light and dark and color to come up with decisions. These are polished, improved, honed until the final product is presented. Here we get a glimpse into that process and the difficulties involved.
Arthur C. Clark stated “Any sufficiently advanced technology is indistinguishable from magic.” (His “Third Law” in Profiles of the Future). What he probably did not imagine is that just a few decades later we would be so blasé and jaded about this “magic” all around us that we now expect it and complain when it does not do quite as we imagine.
This will be an awesome product!