Carl from United States  [18 posts]
12 year
Any idea why the object recognition module would see a nonexistant object in an empty background space?  Here's a screenshot of what I'm talking about...

Anonymous 12 year

Any chance you can post the original template and the image above without any annotations (i.e. the source image)? Also, what recognition method are you using? Feature points? Shape?

We can then check why you are getting that ghost image.

Carl from United States  [18 posts] 12 year

Sadly, I no longer have the original source image.  I was using Haar
recognition at the time, attached are the .haar file and the .robo file
of exactly what I was doing.  The background isn't filtered, I used a
black surface in order to try and keep things simple.  I'd been having
difficulty having it recognize it consistantly when rotating it on the
plane of the surface, I gather training is properly accomplished by
rotating the object into several positions and hitting "Add Object"
and specifying an identical name each time, although this appears
to result in multiple objects rather than a single object recognizable
across a full circle of rotation.  Originally I'd tried using Feature Points
recognition but RoboRealm began behaving strangely, claiming not
to be able to load the image and giving a gibberish in the error popup.
Once it started that I couldn't get it to behave until I uninstalled and
reinstalled RoboRealm, then I switched to Harr recognition. I notice
now when I open RoboRealm the path looks strange as displayed,
looks like "œyirogram F¨yi\Objects\" rather than the actual path.

Thanks for your prompt attention,
Carl from United States  [18 posts] 12 year
Files didn't come through on the first attempt, hadda zip the harr file first heh :)


Anonymous 12 year

Thanks, that was helpful. I think we found the problem and have uploaded a fix for you to test. v 2.44.27 has this update. Can you download again (use the Download button in the Options Dialog) and see if that helps to remove the error.

Carl from United States  [18 posts] 12 year

Thanks, it's much more consistant across the rotation even with only
a single train, although the bounding box varies considerably.  It still
produces the ghost at times though.  And another issue I forgot to
mention is that the OBJECT_ORIENTATION variable doesn't update,
remaining at some arbitrary value.  I've supplied the files for this
test, including the source image this time.

Thanks again :)


Carl from United States  [18 posts] 12 year
Although it does misidentify, possibly since my methodology
is simplistic at this point.  I don't think my finger looks that
much like a lighter :)

Anonymous 12 year

It doesn't but that's why the system rates it as 75% with the ghost image being 70% and both lighter images being above 95%. So if you adjust the min threshold to 80 or 90 that will eliminate those matches. It is set to the default 70 right now which is lower than it needs to be just to ensure that something matches initially.

When adjusting that turn off tracking since that has an even lower threshold of 50 otherwise the bad matches will remain.

Give that a try and see if that helps.


This forum thread has been closed due to inactivity (more than 4 months) or number of replies (more than 50 messages). Please start a New Post and enter a new forum thread with the appropriate title.

 New Post   Forum Index