Riya Beta

In the next few weeks we will be releasing Riya into Beta (and hence public access).   Those of you on the alpha waiting list will get it first.  I wanted to take some time to talk about the new version, thank all of your for your feedback to the alpha, show you what we did with it, and ask for even more feedback.  Before we start... a special thanks:  The entire Riya team has been working flat out - 16 hours a day for almost the last 45 days (since the last testing ended to prepare this release).  The team deserves the credit for getting this beta done. Keep scrolling down - the supposedly WYSIWYG typepad editor is anything but what you see is what you get.

Homepage:

• Many people wanted to try Riya before adding their own photos.  Others told us they wanted to find photos of themselves before uploading their own photos.  Done - Now Riya will let you search public Riya photo sets without having to login.  In addition, you can do public searches,  see public albums, see public people, see photos by time, by location, and by text we find in them.  Riya is like any photo search engine for public photos.

The Nav bar

• Folks were confused that our old navigation was not always available on all pages.  Now it is.
• Others wanted the number of photos to be available on all pages.  Now it is.
  
• We've organized the navigation into two sections.  The left is all of the ways you can slice and dice / search your photos.  You can search by people, albums, time, text, etc.  On the right is the workflow of how you upload, train, and share your photos.  Most of the new pages were conceptualized and mocked up by Tara .  Tara is a UI design genius.  She hates it when I say that.  In fact for months she tried to hide her talent but in the end it was just apparent.  I love the new look of the site and Tara is responsible for a lot of it (with lots of help from Uma and Dan).  And she can blog... what a bonus!
  
• One of the things we hate about other sites is how much space the header takes up.  Uma actually spent some time to compact the header so it doesn't interfere with your photo experience.
• Notice the new messages section at the top.  If you do a search today you will see some photos that are greyed out and say "Get" permission on them.  This is where you will recieve the requests from others allowing you to "discover" photos of you that other have.

Search

• A major theme of the feedback we received that we need to allow better sharing and viewing of shared photos.  There are changes throughout the site to support this.  This is the first.  Notice now that you can search and see the result of that search in your photos, your friends photos (if they have linked to you in Riya), and in all public photos in Riya.  With permission, you can search for which of your friends have photos of you or of Tomichi point (where you might be going hiking next week) with one search.  We hope this unleashes a whole new way to get the photos you want.
• A lot of folks wanted our advanced search to be implemented.  We didn't get time to implement all of it, but we focused on the major point which was the ability to narrow your search to given time or date range.

Search Result

• You said: Searches are too slow - fixed.  Most are < 1 second now.  For you Ajax nerds believe it for not we speed this up by doing less Ajax and move more stuff to the server.  Somethings just don't belong on the client.
• You said: I need to be able to sort by time in addition to the default (Riya Rank or relevancy) - Done - Neelesh and Krishna not only speed things up but they added a lot of functionality to our search engine.  One of the key things they did was make our engine multi-cast.  Hence our new speedo is now far more scalable than it was before.  If you don't know what multi-cast really means in detail... don't worry I barely do ... ;-).
• You said: Give me an RSS feed of this search - Done
• We added the ability to increase the number of thumbnails on a page in increments of 50 (not shown) using Ajax (so no need to refresh the page)
• A second sharing change is that you can now just cut and paste the URL of the search shown in the browser address bar (not shown in this screen shot) and email it to a friend and they will get the same search result. 
• You said: Let me tag all of the photos (not shown above).  Now you can control click a few of the photos (or all) and tag them in bulk.

Detail view of each photo

• You said: Let me comment on a photo - Done
• You said: Give me a slide show - Done
• What are those gray boxes - I'll leave that suspense to the release - it is very cool - Of course my man Dan Chaio did it.  He even added little touches like when you hit next it automatically scrolls down to push the header off the screen and show you the photo full in the screen (in 1024 x 768 resolution).  Dan and Uma also added the ability to use the right and left arrow on your keyboard to move through the photos without having to click on the arrows with the mouse.
• One of the hardest decisions we made was to remove the long scrolling photo detail pages.  We used to call this Korean View.  Why - it was in Korean 1.5 years ago when I first saw the future of the broadband web that I saw many web pages that scrolled forever vertically.  Sometimes these pages were 10 MBs big ... but with broadband it didn't matter.  Instead it was more important to save the user a click to see the next photo.  We tried to implement this every which way, but finally we just couldn't get around browser limitations and compatibility issues.  Dan has probably reimplemented this numerous times.  So no more Korean view....;-(.
• When tagging a face the Riya system used to prompt you to give it more photos of the person in an annoying nag dialog box.  You asked us to remove it.  We didn't because we need the extra examples to do better, but if you have already give us enough it no longer asks you for more.

People Tab

• A few UI changes on this page to make it more clear.  Another link to share all of the photos of that person with them
• If you click on the friends photos link you can see all of the people your friends have trained.
• Piyush did a great job on the whole addressbook infrastructure.  This code was a mess before and wasn't scaling if you had more than a few people in the addressbook. He fixed it and with Uma they have made it just plain beautiful.
• You might not be able to see it, but look closely  and see the check box that lets you give the person here access to search your library (linking to you).  This is the key to all of the friends searches.

Albums Tab

• No real changes to this page except if you click on the friends link you can see the albums your friends have shared with you and let you have access to.

Time Tab

• See what dates you have taken photos on (highlighted in green) and click on the date to see the photos from that day. 
• See what dates your friends have takent photos and click to see
• See what dates public photos exist in Riya and click to see those dates.

Location Tab

• Okay this is hot.  Krishna analyzed the text we read in your photos, text in your files names, and text in your folder names to find known locations.  For example if he found Las Vegas Hilton, he looks up the address, then the latitude and longitude, and finally puts that photo on the map.  Good work Krishna.
• You can also see this for your friends and public photos (we choose the top 200 randomly).
• Yes our first mashup...;-) I guess we are Web 2.0 now.

Text Tab

• UI same as before - We improved the handling of special characters like dashes and parenthesis.  Now we can read phone numbers in photos more accurately.
• We didn't improve much else in text because we asked Vincent, our text recognition star to work on some of the face re-engineering for the last 45 days (see below).

Add Photos

• Our uploader was one of the most problematic areas of the service in the last alpha test.  60% of all users had problems with it.  Many complained about it hanging, taking too long, and not detecting enough faces (which occurs in the client as we upload). 
• Nitin and Vinet reworked the entire uploader to make it more stable and usable.
• Azhar the super planner and driver of our development process, hired Stag QA consulting in Bangalore to install and test the uploader in almost every configuration we could think of (every version of windows, with this popup blocker, with this anti-virus, after we pulled the network connection, etc).  This was a great call on Azhar's part.  I hope we caught everything. 
• So far I'm been talking about cool UI features.  The core of Riya is the face recognition and detection.  A few people who reviewed the product said that Riya not detecting most of the faces was a show stopper for them (I'm trying to remember who wrote this on this blog - but we got it a few times). 80% of our effort over the last 45 days has been spent here.  We've completely revamped the  recognition and detection engines.  Diem, Danny, and Vincent completely retrained our face detectors in a mere 45 days - an amazing feat.  Hats off to them - the results are so much better - in some of my photos 2x more faces are found.  Danny turned every machine in our office into a massive compute grid of 30 machines to speed up the retraining by weeks.  Now Riya gets more profile shots and more tilted heads than it got before.  We are nearing 75% of all faces detected vs. around 55% before.  The detection system used to be just average (our focus was more on recognition), now our detection system is state of the art.  We still have a little ways to go (which we will do later - We still can't find a face is we can't see both eyes in the photo.
• Speed - Here we didn't improve at all.  In fact, in order to get more faces, the time it takes to process each photos has actually risen by almost 3x to 30 seconds a photo for 5MP photos.  Many people were wondering why our uploader took so long.  Actually it isn't uploading most of the time, instead it is looking for faces in your photos and isolating the coordinates for each.  The UI of our uploader highlights this a little to explain what it is doing all this time (not just uploading).

Training

• As I said earlier face recognition and detection is where we spent most of our effort.  A descent minority of the users of the first alpha felt the recognition should improve before we go beta.  We took this to heart.  We have worked on this the most and caused the gap between releases.   Kuang-Chih, Baris Smengen (A great similiarity researcher who almost became employee #2 but finally joined us just now), Lorenzo, and Drago have worked 28 days in a row (including weekends) to improve.  God bless them - we told them at least recharge and they didn't want to - preferring to finish this release.  Their work has been on the core recognition algorithms.  We are still waiting to measure the improvements they have delivered.
• We completely revamped our training wizards and UI.  Burak came up with a super smart idea whereby a user has to only train us on one face of a person and we return the maximal results.  If you remember before you had to train us on many of a person in what we called "bulk training".
• Dan then built a new very Ajax interface where you can quickly train by dragging a name to a face or by ctrl clicking and dragging a face or a few faces to a name.  Lastly you can  just tab through and quickly train a bunch of faces without ever moving to the mouse.
• Nikhil and Sandeep built a whole new recognition framework to handle this new recognition system.  Even the actual mathematical vector we are using to represent the face has changed so we will need to delete all photos in the current Riya system to prevent backward compatibility issues.  The efficiency and leadership of these two guys continue to blow me way.
• Once done the recognition algorithm runs and then returns the results for you to review before commiting them.  Previously it just applied the tags, now as a user you have more control.  You X out the mistakes and continue.
• Look at the results for Tara - they are just amazing.  Even I wouldn't have thought most of these were Tara but they are.  Once a user confirms all of these photos become part of the training set improving recognition even further.  Ignore the formatting errors - we are fixing those.
• This is the most risky change we've made - but necessary to improve the results.
• Folks told us they wanted better recognition - we have built a new engine - we'll see if it is good enough.
• Folks told us they wanted more instant recognition - before they didn't know if recognition had run yet or not, now it is a wizard.
• Folks wanted to label a few photos of a person and get lots of results back... now we are closer to that ideal.
• We needed more examples of each person in the training set ... now we get hundreds of user confirmed examples.
• One thing we don't know, is the performance impact of these changes.  It used to take 30 seconds to run recognition (not detection) on large sets of 5000 photos.  We have not yet benchmarked the new system.  We will know more in about a week.

Auto-Train

• One of the key things we learned in the Alpha was that people really needed the help of their friends to build a good training set of those friends.  Most of us  don't have enough photos of our casual friends to build a robust training set of that person.
• This is one of the reasons Riya is a web based service to allow this critical share.
• The new auto-train tab makes this easier by analyzing your address book for people you know that Riya already knows.
• Of course the first 1000 users of Riya will probably not benefit much from auto-train because there won't be many other people in the system.
• Our India Research team also came up with something very smart that will boost recognition significantly. I will let you discover this once you use the beta.

What we didn't do

There were a number of pieces of input we didn't get done.  I'll list them here to at least tell  you all that we heard them (email me if I missed any of them).

• Mac uploader - we tried to get this done and still might make it for the release, but we had a hell of a time finding a consultant who had the expertise to write the OSX application.  We finally found someone but he couldn't work more than 20 hours a week.  We are still working on this.  Tanvir - a rockstar engineer who is joining us on the 1st is seeing if he can accerate this.
• Speed of uploader - I already talked about this ... we  thought it is better for Riya to be more accurate right now than fast but inaccurate.
• Blog this and Rotate - We just didn't get to it.  No reason we shouldn't do it.
• Flickr importer and exporter.  The Flickr API is available for this but you can't use it for commercial purposes and once Riya ads advertising, it will be a commercial service.  We will approach Stewart Butterfield and Yahoo about using this API for commercial purposes but we haven't gotten it done yet.
• Full stand alone client for running Riya without a network connection.  This was a strategic decision.  We couldn't figure out why users would use another client and switch from Picasa/iPhoto/Adobe and how you really got a community (friend search) going standalone.  I know uploading is a pain but it the social web is very powerful and yet can still maintain privacy.  Of all the decisions, I'm the most confused about this one.  Any input is welcome.

SoBetaItHurts

• Riya is still very experimental
• Riya beta is not generalized recognition system.   It is specific recognition system that works well on your family photos taken with a flash, with high resolution, with time stamp in each photo, with people smiling and looking at the camera, and with events in which you have lots of photos of each person.  Please remember this.  I call this the Ferrari offroad problem.  Riya is works well on nice paved roads with losts of straight aways and no cops...;-).  Not dirt road (low res photos from the web with lots of unique people, etc).  It will not work for celebrities photos you download off the web.  It will not work for DMV style photos in which the clothing is not present. 
• Riya is only as smart as a 2 year old (or maybe even 1.5 years old).  Deven is about this age and he correctly identifies grandma most of the time but every now and then he needs help.  To learn a new person he needs lots of repition to get it.  If he hasn't seen them in a while he forgets, etc.  Riya is like a 1.5 to 2 year old kid.  That being said, it is better than tagging your photos by hand....;-).  Riya will get better over time, but it will take time.  The opposite is also true, Riya can't do something that a human can't do.  If I show you the photos of all 10,000 people in a  stadium and let you study for even a week and then show you photos of people in the stadium and ask you to identify them. You will make more mistakes then you will get right.  If a human being can't do it Riya can't do it.  Riya works best for 400 or so unique people.

Coming Soon - Riya Beta

Send an email to alpha at riya.com if you want to get it early.  Send me all of your feedback by posting here or by emailing me at feedback at riya.com .  We are so excited and anxious about this release and look forward to you input.

Comments

[Abhishek Goyal]

Tagging is the feature i was really looking forward to. It helps a lot when you want to retrieve photo or rather browse photos.

While trying the public search, i feel the thumbnails created are distorting the image a bit (may be bad resolution of original can be the reason), but thumbnails (on the search result page) doesn't look too neat.

The search queries are pretty fast now, but the thing is the thumbnails takes almost a minute to load (in firefox) and so in the end i get to click or scroll (14'' monitor) results only after a min or more (256Kbps connection).

Congrats on getting sajan in team, i heard he has become quite an ajax expert now :-)

[Halai]

this looks brilliant guys. good luck. the UI seriously needed an overhaul in more places than one. rest assured, the screenshots are amazing. rather looking forward to the release.

[SamD]

Thank you very much from all the previews and laying out *exactly* what is happening with the Riya Beta Release. I am really looking forward to getting in there later this week. Thank the team also, by the sound of everything they absolutely rock.

(Also: Flickr Im/Export should be high or your "to-do-after-first-public-beta" list.)

[henri]

I see a lot of improvement since I tested the riya application last year during the alpha test and since I was, two weeks ago, a member of the Ground truth test (a hard job too, because tagging was done without auto recognition).

I'm happy to test this new release in the next weeks

[haifischjunge]

For your strategic decision not to built a stand alone client i would agree, however there should be a solution to connect with the the usual picture organizers. riya should be able to syncronize fotos, albums and tags with the metadata of the most important applications (Picasa, iPhoto, iViewMediaPro ...).

Just speaking for myself I have all my digital pictures stored on an external 2,5" harddrive and organized with iviewmediapro - keywords written direct to the files with iptc - and I would like it very much, if could get also all the tags from riya into iview or direct into the files.

I think your standalone app should not organize pictures, rather sync metadata and pictures with riya.

[Ruby Sahiwal]

I'd like to use auto-categorization of my photos within iphoto without downloading all my photos on to your website. Is that or will it be possible? Great concept BTW. - rs