|
New Object Recognition Module
Danh from United States [79 posts] |
13 year
|
Steve,
Wow! Thanks for the new Object Recognition Module. I've been playing around with it to recognize playing cards. There's a lot of experimentation ahead for me - forget sleep.
This latest module just took RoboRealm to the next level IMHO. Thanks!
Attached are some of my preliminary tests.
Danh
|
|
|
|
[29 posts] |
13 year
|
STeven,
How does this differ from the AVM module?
|
|
|
|
Danh from United States [79 posts] |
13 year
|
Steve,
I've been testing out the new Objection Recognition Module. Here are some bug reports:
1) When I hit the "Add Object" button the .png file that is created appears blank if I select the region by first moving my mouse to some random image location and then press shift and mouse drag to select an area.
2) The following files crashes RoboRealm if I try to train using Haar (it's the 52 cards in a deck). The other methods don't crash (feature points, shape, cross correlation). BTW- I found that Haar recognition works best for recognizing the card values. Here is a link to the files:
http://www.imaginerobots.com/all_cards.zip
Thanks,
Danh
|
|
|
|
Anonymous |
13 year
|
Danh,
Yikes! I wouldn't pack for Vegas just yet. :-)
Thanks for the errors. Both are now fixed in the current version. We've also improved the Haar and Shape matching to improve the results. Hopefully the cards get closer to their actual matches. We hadn't dealt much with very similar templates such as a deck of cards.
DM,
These are similar yet different from AVM. Basically this module has a couple more techniques that provide similar results to the AVM but using different techniques. They are all object matching but because of the way these methods work some will work better for some objects. That's why we created this module ... to allow for multiple techniques to be included for testing ... the AVM being one of them albeit in its own module. The AVM may share some of the included techniques but as it uses multiple sample images (you'll note that during the training phase it is sampling many images from the webcam) it is creating a different type of model that what is included in the Object Recognition template. Basically the more techniques the better!
STeven.
|
|
|
|
Danh from United States [79 posts] |
13 year
|
Soon I'll have my poker playing robot and Vegas here we come!
Steven, thanks for the updates. Here are some more observations:
- In the Filter Matching panel when I change the "Select Top X matches" box to some value it doesn't get reloaded the next time I start Roborealm
- For the "Filter Orientation" min angle is it possible to set it to a negative value? This would represent the real world position of a card that is slightly rotated clockwise or counterclockwise. I found that a max angle setting of 360 degrees worked but it takes up a lot of processing power. I really only need like +/- 25 degrees for a playing card
-I've attached an image showing a new scheme I have for detecting cards quickly. These are special Super Jumbo cards from Hoyle. There's a training template for the Suite and then another training template for the numbers. Having the separate templates allows me more flexibility in tuning the confidence level, filter scale, etc. Preliminary tests show good results. The goal will be 100% error free detection. I think using color will also help too. Any thoughts on this approach?
This fine tuning of the image recognition is really helpful and makes all the difference in the card detection capabilities.
Thanks,
Danh
|
|
This Thread
