I started testing
Riya, the much-talked-about face recognition company, and here are my notes so far:
- There are people I don't care about having Riya recognize -- I'd like for them to NOT show up in the manual training. There's isn't a way to tell it to stop showing me certain faces so at this point, when I go into manual training, the first three or four rows are full of poeple that I just automatically skip over. And this makes each training "session" less useful.
- The first thing I wanted to do after I did some "training" was see what faces Riya had properly recognized based on my training. The way I see it, face recognition is the reason why people are going to use Riya so it's the place a user should be taken after they've done the work of uploading and "training" (instant gratification). As it is, I had to hunt and peck for this screen, I wasn't automatically taken there.
- On the "People" tab, I want to have a count of how many photographs it's found for each person I've trained it for... This would be a good measure for me to quickly get an idea of how well it's done.
- On the "People" tab, the left and right browser arrows don't appear to work in the "Browse unrecognized faces in your photos" section.
- I'm curious about the percentage of faces (in my photographs) that Riya has recognized on its own. I have the total number of faces recognized, but I don't know how many of these were because I specifically said this face = this person. On one end, this number could be zero (this is the scenario where I've taught Riya 100% of faces in my photographs) and Riya would be adding no value. On the other end, It could be something high and that would indicate (assuming a low rate of false positives) that Riya's recognizing a lot of faces.
- The uploader wasn't as effective as it needs to be -- it should have had Picasa-style directory selection/de-selection. As it is, I do a lot of exports from Picasa and so I have lower-res duplicates of a lot of my images at My Pictures/Picasa Exports/[Album Name]. And all of my native resolution pictures are at My Pictures/[Album Name]. So as a result, a believe Riya's uploading a lot of stuff twice.
- I manage all of my photographs using Picasa. It would be nice if Riya integrated with this so that it picked up edits / crops / rotations that I have made to photographs (Picasa doesn't change any of the source image files, it just stores the changes in an INI file and applies them whenever you try to do anything with the file outside of Picasa like send it out for a print, do an export, email photographs, etc.). If there was integration, I could also tell Riya to only worry about the photographs that are marked with yellow stars.
- Is Riya uploading the high-resolution images (3-4MB per image)? Or is it just doing image processing locally on my hard drive (utilizing the higher-resolution data) and then uploading a lower resolution version of the image along with the results of that image processing to Riya? Does image recognition quality get incrementally better at higher resolutions and then plateau past a certain resolution?
- Summary: So far, it's been fun and interesting. But what I really need to do is assume that Riya will integrate nicely into my photo ecosystem (ie my multiple cameras, my portable hard drives full of photographs, Picasa, and flickr) and figure out whether this facial recognition stuff is actually
useful. It's certainly cool, but why do I need to have photos referenced by person? Does it improve the photo sharing experience? It's more often that I want to share photographs from an event or by theme, but it's not too often that the theme is a person. I can think of some instances where it would be useful, like if I'm putting together a slideshow for someone's birthday -- then having all the photographs of that person is useful. If I'm looking at really old photographs, I think it might be useful to be able to pluck out photographs by who is in them. Maybe. Like I said, this is one that I have to figure out.
1600+ photographs uploaded so far, I'll continue to report on how well the technology works and, ultimately, how useful I find it.
Tag:
riya
* Adobe's office in Delhi is really nice -- a big thanks to them for giving us their conference rooms, for feeding us lunch, and last, but certainly not least, for the fast and free wi-fi. And while I'm at it, big thanks also goes to the folks that catalyzed and planned the event: Gaurav, Jonathan Boutelle, Manik Juneja, and Amit Ranjan.
* Want to befriend an Indian engineer / technologist? Buy him a beer. In particular, buy him a Foster's. :-)
* Some of the presenters were talented speakers -- 
