Log in or register to post comments

More insight on trackables? Should they be snapshots of something *printed*?

December 21, 2010 - 2:57am #1

Here is one example of a 4-star trackable that seems to track worse than a 0-star trackable...

I haven't printed this out yet, but I generally have luck shining my evo cam at the screen to trigger a trackable.

For some trackables, I have to print them out, take a snapshot, and upload the snapshot as the config.xml trackable. I am wondering if that's the case here, or if there are other reasons why this wouldn't be a good trackable. (But, why is it showing up as a 4-star trackable?)

Here is the trackable http://uploadbox.com/files/278a18b517/

http://uploadbox.com/files/278a18b517/

Re: More insight on trackables? Should they be snapshots of some

December 22, 2010 - 6:16am #4

Right, this tracker only works with known images, and is expecting the precise image that you ran through the preprocessing step. It isn't going to work for tracking 3D objects in the real world (such as your pile of wood, or even your keyboard). It also can't track random reconfigurations of the known image. Those are different tracking problems altogether.

- Kim

Re: More insight on trackables? Should they be snapshots of some

December 22, 2010 - 12:40am #3

For the ImageTargets - it seems like you have to use the print-out or screen-readout, and that, for stone or woodchips, you can't just point the user to the location: For example, if you just shine it at a pile of wood that looks very similar to the imagetarget, but isn't the same segment, it won't work. (A quicker way to check this would be to photoshop flip the original imagetarget - no detection.)

For ImageTargets: something like scanning that sign on a building in Rokus Reward would work for sure, but if you graffiti-ize it a bit, it'd likely not.

So, if you could just track -- say -- a random pile of stone or woodchips (rather than that particular shot), you'd be able to automatically spawn unique "scenic design" each time you play it. I suppose something like that would be more suitable for the Amazon-Turk-based image recognition app (oMoby on the iPhone).

If imagetargets could recognize similarlty, then... A fun iPhone-esque app would be to point your device at a field of grass (any field of grass) and instantly see virtual flowers and bees pop up, and maybe play some sort of flowers+bee game o.O -- similarly, point your device at the sky, and instantly see virtual balloons to pop! or UFO's to shoot!

Re: More insight on trackables? Should they be snapshots of some

December 21, 2010 - 6:08am #2

The rating system isn't perfect, and it looks like you've stumbled upon one case where it doesn't work very well. That image seems like a good trackable at first glance, it has a nice even spread of features. The problem is that the pattern is too regular. Each square key looks like the next, and the letters on the keys are too small to make good features. The tracker will have a difficult time differentiating one section of the keyboard from the other.

In general, patterned images make poor trackables. See the Create Image Targets section of the Dev Guide for more guidelines:

https://ar.qualcomm.com/qdevnet/developer_guide/345

Finally, you should never have to take a snapshot of a printed target to use as input to the My Trackables system. It is important that the image uploaded is a clean 2D representation of the image that will be printed, and if you take a snapshot you risk introducing strange angles and offsets, not to mention camera warp. That might make it a little easier to track with a particular device held at a particular angle, but makes tracking much less robust across all devices and orientations in general.

- Kim

Log in or register to post comments