12

Bib Recognition for Sports Photographers (Read 1553 times)

    Off the beaten path, but still related to our joy, running.

     

    I've been working on a program to help sports photographers tag bibs automatically. I'm still working on it, but have a demo up here: http://bibmatch.com

     

    I'd love to hear any feedback, good or horrible -- particularly from someone who's a sports photographer.

    Creator of RunForth@robraux | shodless.com


    Interval Junkie --Nobby

      Pretty slick.  I'm not a potential customer (I don't do sports-photos, just street), but if I were, I'd like to read about the estimated accuracy of your process.  Estimated percent of false-positive, or missed tags.  Is there any way that the software indicates that it isn't sure of a number?  Or that there was a bib in the photo, but it couldn't read the bib?  A simple green, yellow, red confidence level associated with each of the bibs would be nice.  This would allow the photog to manually address anything the processing thought it missed; a missed bib is potentially a missed sale.

       

      The tool itself looks very simple from the demo; nice work.

      2014 Goals:  sub-3 Marathon 

      Current Status 08/28: Slowly working back up from a pelvic stress fracture.  4mil distance PR w00t!

        Thanks for the feedback. I'm planning on providing statistical information like that (not ready yet). The confidence level information is a wonderful idea, thank you!

        Creator of RunForth@robraux | shodless.com


        Interval Junkie --Nobby

          Also, something you should address in your advert - how does the processor handle when a digit of the bib is covered by an arm or something?  Does a 1234 become 234?

           

          This is something the confidence indicator should give a "yellow" to if the processor is able to detect the situation at all.

           

          Good luck with your product!

          2014 Goals:  sub-3 Marathon 

          Current Status 08/28: Slowly working back up from a pelvic stress fracture.  4mil distance PR w00t!

            That's very cool, Rob.  To expand on stadjak's suggestion, maybe having a threshold setting that uses the confidence level to decide whether or not to tag the photo with the detected number.

             

            Extending this to videos would be a whole 'nother ball of wax wouldn't it?

            Well at least someone here is making relevance to the subject.

              Extending this to videos would be a whole 'nother ball of wax wouldn't it?

               

              Yes, and no. There's a lot of computational wizardry (read: math) going on behind the scenes regarding computer vision. If the algorithms can work for images, there is a good possibility to extend them towards video as well.

               

              The website is an attempt to gauge interest and solicit interesting feedback like you and stadjak have posted.

               

              There are a lot of options regarding partial bibs, or a shot of a person w/o the bib. That's were things get more and more complicated... all of the edge cases. Coupling the bib recognition (this is much more than just OCR) methods, with facial recognition, yields very promising results.

               

              I want to know that there's a worthwhile interest in it before I pour in some more serious effort. It is fun to work on though!

               


               


              Creator of RunForth@robraux | shodless.com


              day after day sameness

                I would guess (key word) the folks who do the systems police use which read license plates in real time have a lot of the same issues around partial hits. One suggestion to throw out is to consider how to handle bibs on crooked or sideways. Had a race last year where the timing tags were mounted on the bibs wrong, and everyone ran with bibs pinned sideways.

                Choosing my words carefully has never been my strength I've been known to be vague and often pointless

                  MT, photo-red type systems use an actual human to verify license plate matches before anything is sent out.  I can't speak for every state, but I know the two I looked into both required that human check.

                  “Everything you need is already inside.” -- Bill Bowerman

                    Small site-related glitch: I can't read your privacy policy in Chrome because it appears beneath the video.  (But it works fine in Firefox.)

                    "Because in the end, you won't remember the time you spent working in the office or mowing your lawn.  Climb that goddamn mountain."

                    Jack Kerouac

                      Overall, looks sweet.  Great idea, if you can prove the statistics, I'm sure you'll have a lot customers Smile

                        Thanks for all the replies guys. I appreciate each one.

                         

                        The sheer number of "edge cases", where the system needs to be smart is both daunting and exciting. All of it's doable though, which is the fun part.

                         

                        @Jehu - thanks for the heads up on the glitch

                         

                        I welcome any other feedback you may want to throw my way, and if you know any sports photogs, please pass it on. The more of those folks I can talk to, the better!

                        Creator of RunForth@robraux | shodless.com


                        A Dance with Monkeys

                          Don't all the major running photo companies, such as marathonphoto, already use something like this?

                           

                          Anyway. Here is what I would love, as an RD. I have a stack of photos and the database with race numbers and email addresses. I'd love to have the program ID the photo and match it to the email address, then email the photo (or a message) to the runner.

                          Gig


                            Great idea. I forwarded the link to Annapolis Striders.

                              Don't all the major running photo companies, such as marathonphoto, already use something like this?

                               

                              Anyway. Here is what I would love, as an RD. I have a stack of photos and the database with race numbers and email addresses. I'd love to have the program ID the photo and match it to the email address, then email the photo (or a message) to the runner.

                              Trent, that's what I'm trying to figure out, however the answer from everyone i've spoken to seems to be: "they correlate photograph timestamps with when a bib crossed a check point.

                               

                              This works decently for large races, with multiple checkpoints. Anything without that big time setup, or any pictures taken in between the checkpoints have to be done manually.

                               

                              Once the technology is in place, there are a lot of great options for it, just like you mentioned. There's even the possibility of having a photobooth at a race where people search for their bib# right after the race and print the photos an hour after they've finished. 

                               

                              Everyone seems to be looking for tech like this, but no one seems to be able to provide it, because the barrier to entry is high, and perhaps because the market just isn't there. That last part is what I'm most interested in. 

                               

                              So thank you both for the feedback, I appreciate you sharing your thoughts. 

                              Creator of RunForth@robraux | shodless.com


                              Blaine Moore (MM#2867)

                                Very cool, definitely interested in following developments. Signed up for the email list.

                                Run to Win
                                24 Marathons, 17 Ultras, 16 States (Full List)



                                12