I’ve finally uploaded the Markerless AR stuff I’ve been working on for the past few months or so. Still more work to be done. At the moment it’s just meant to be a demo application that you can play around with. I don’t have any immediate plans to polish it up and package it as an API or anything. You can download it from this page.
Haven’t posted up anything for the past weeks. Been busy working on this Markerless AR demo code, which should hopefully be released soon.
Recently. there have been a lot of interest in getting AR working on mobile phones (dubbed MAR for Mobile Augmented Reality). The results from papers I’ve found have been quite impressive. But I thought it would be just as fun going the other way and trying to get it running as fast as possible on a powerful PC. This allows the user to do more intensive processing and yet run in real-time. Complex 3D models, animation, particle effects, sound, and maybe even some spare cycles for a frame of Crysis … thats the idea anyway.
To do this I’ve been using the GPU where possible. So far the GPU is used to: extract patch features, build the feature descriptor, match features using a simple K-Tree structure on the GPU.
I’ve spent some time trying to put the Homography calculation on the GPU with poor results, probably because my card does not support double types, which may be necessary for SVD calculations. My laptop’s graphics card is a Nvidia GTS 360M. I can get 20-30 fps easily. I’d imagine with the latest and greatest Nvidia card on a desktop PC, it can only go higher.
In the mean time, here’s a recent screenshot. The room was poorly lit so the webcam ran only like 5 fps. But the processing, at the current time, was capable of 33 fps or so. I’m hoping to increase that with some further GPU optimisation. I’d also like to point out it’s 33 fps WITHOUT any ‘real’ tracking code. I’ve only used a bounding box from previous matched frames to limit the feature search. It’s pretty much running in semi-detection mode all the time. One of the main reason why I didn’t put in a tracker yet is because my room is poorly lit, which makes the webcam very susceptible to blurring. The blurring will easily kill any tracker. So it would be kinda moot to test.
I also got rid of the teapot and replaced it with a 3D model of Tux! which turned out to be a pain in the !@#$ a$$ to get up and running, but I’ll leave that rant for another post.