A lot of the other CV based AR approaches that I'm familiar w/ rely on a distinct border region to the target / marker. QCAR's image tracking seems to be able to handle 'borderless' images pretty well, provided that there's adequate contrast and differentiation w/ the backdrop. But does the tracking algorithm make any assumptions about the border region of the image? What are best practices concerning the continuity and geometry of the border?
- Sort Posts
- 4 replies
- Last post
Re: Role of the image border?
That's right, repeating patterns won't track well in general, the tracker can't distinguish one section of the pattern from another. The features should be somewhat random and noisy, but still distributed across the target.
Feel free to send targets to
if you feel they should have worked well but didn't. We'd be happy to review.
- Kim
Re: Role of the image border?
Interesting. Thanks.
Just make sure that the portion of the image the user will focus on has a nice, even distribution of feature points.
Yes, but not too regular it seems. I've gotten false positives from the TMS on grid patterns, for instance. They'll rate a 5 but are never recognized.
Re: Role of the image border?
The outermost border is not actually used in tracking, instead what matters is the distribution of features across the target as a whole. It is certainly possible to create non-rectangular targets where the features are concentrated in a portion of the rectangular image (e.g. a circular target). Note that these targets might get a lower score from the TMS, but experimentation should reveal whether the target performs to your needs. Just make sure that the portion of the image the user will focus on has a nice, even distribution of feature points. You can click on any target in the TMS to see the location of feature points across the target.
- Kim
The false positives that I've encountered so far are predictable, when you recognize how the tracking operates. If you variegate the image a bit, by using a complex border or islands in the grid, they can perform well. It's just pure grids that are an issue.