astatic notes

Looks like I completely forgot about my blog. But if u follow me on twitter u know what I was doing all this time. Writing blog posts require lots of time while posting to twitter is really fast Let me tell u whats going on around.
FITC SFO
I was in SFO doing a session at FITC, besides the fact that I was pretty nervous cause of my first serious session it went pretty well I guess. Lots of people were interested in my subject and I got lots of comments after it. But I know that next time I should prepare better.
I was very happy to meet lots of online friends in person. the first person I met was Mario and the first thing he said was “c’mon, lets have a dinner with us!” skipping the fact that I’m not Asian food fan it was fun to see all this guys together talking about random things. Thank you for that! I enjoyed the whole event and what I like most of all it is the chance to meet all that people in person, to speak to them, to discuss and to share ideas. Special thanks to Bartek and Rob for making me a company when I didn’t know what to do or where to go. Greetings to Alan Shaw and Joa for hanging with me during lunch breaks.

NO MORE SURF
After SFO I continued ASSURF development but focused on optimizing main algorithms. The heart of ASSURF was keypoints descriptor adopted from well known SURF algorithm. I always knew it wasn’t good idea to use it because it is patented with all consequences. My attempt was to create fast and strong descriptor that can be used in any project without caring of legal issues. I wont go in details here but I’ve ended up with random tests based descriptor that stores point information in bits. As the result it is very fast in computation and blazingly fast to compare using simple XOR (^) operator.




Next phase was to optimize after detection step aka tracking. My first attempts was to use different algos (Optical Flow, SSD) to recover previous frame points and to skip new frame computations. But that works very unstable. I chose different approach – template based tracking. It uses precomputed reference image matrix to find object bounds in provided ROI. This approach requires training process but this can be done offline and then loaded into the tracker. Tracking step is very fast it takes about 1-3ms to track single object and no point detection/description/matching needed.

And finally we need to estimate 3D pose of the detected/tracked object. This can be done using Homography matrix that we already estimated during tracking step (see green border around image). I shared the class that allows u to decompose Homography into Rotation matrix and Translation vector in 3D space (see Homography decomposition code) It uses SVD routines for matrix decomposition that can be found in my small Linear Algebra library (I will add Wiki page soon with usage summary)




Besides the ability to detect and track single reference library would be able to handle multiple reference objects. Of course it will be slower in detection/description step but still blazingly fast when all needed references found.





You can find more demos at my Vimeo Album page
As u see lots of changes were done and no more SURF descriptor used so that was the reason to change project name to different one, again. Next project reincarnation will be under ASFEAT name. Also I’m going to change license to GNU GENERAL PUBLIC LICENSE v.2 that means u wont be able to use it for commercial projects without contacting me and applying for different rights. But I’ll also provide open source version of library that will allow u to create open source or non commercial projects. For those who are interested in commercial developments new name will be created:

I’m currently finishing ASFEAT development and preparing public SDK. Expect it to be released in January 2011. Thats all I can share with u at the moment, it is very short summary of what I’m in and what we all can expect from it. If anyone have any questions/suggestions feel free to contact me.