Comments on: Reuters tests facial recognition video search

By: baker

baker — Wed, 19 Sep 2007 17:08:59 +0000

in all likelihood, they’re balancing this “facial recognition” parlor trick stuff with some good ol’ fashion metadata crawling and voice-to-text transcription. and without any data that i can point to for support, i bet audio can provide a way better ID print of a person than some facial recognition software. in fact, that’s probably the last determining factor that constitutes a match. in the case of a musician, it’s probably more like:

1. Metadata crawl for relevant info
2. Speech to text; index and search against lyrics to determine if this is performance or if keywords indicate this video is about or featuring the musician
3. Audio sampling to compare voice print to database of voice print samples
4. Image relevancy scan to determine if the video sections with the highest probability of the musician appearing (based on 2,3 above) meet some criteria to confirm that yes, they’re 70% sure this person actually appears in the video you’re watching.

it’s like magic, except a lot less magical, and way cooler when you call it “facial recognition” technology.

By: Safran

Safran — Wed, 19 Sep 2007 02:27:44 +0000

In fairness, a TRS-80 could recognize Amy Winehouse.

By: discreet_chaos

discreet_chaos — Tue, 18 Sep 2007 22:58:53 +0000

Neat idea, but still a little buggy.

As my first attempt, I typed Jewel hoping for video of Jewel Kitchner, the singer from Homer Alaska. The autocomplete dropped-down Richard Jewell, Jewel Sansom and the target of my search (sans last name) and no matter if I left plain-old Jewel in the searchbox or if I selected her name, Richard Jewell was the one that would always come up.

Sticking with just one-named people, I tried Cher, but they possibly don’t have any video and they had Elvis listed with his first and last. They also didn’t have any Bruce Springsteen, but I didn’t try any newsmakers. I’ll have to give them a whirl later.