|
Object Recognition Speed
Trop from Switzerland [6 posts] |
12 year
|
I am trying to use the object recognition module with the Haar classifiers but I am having problems with it running very slowly.
I am running windows 7, on a Core 2 Duo 2.4Ghz with 4GB Ram
I am using 130 Training images, my roborealm file consists of a screen capture (300x300 area) which is capturing an area or a flash application running on my browser, and object recognition module. Is there something that I am doing wrong/ any tips about getting object recognition to run faster?
I appreciate any help
|
|
|
|
Trop from Switzerland [6 posts] |
12 year
|
|
And when i mean slow, i mean taking about 20-30 seconds to show the boxes of detected regions per frame, its almost as if roborealm has frozen, but it hasnt.
|
|
|
|
Anonymous |
12 year
|
Trop,
That's quite a few images to be recognizing but there are tricks that can be done to speed things up.
In order to determine the best way to speed it up can you include a couple of the 130 template images and a couple of the actual to be recognized frames (from the flash app) so that we can test to see what may work best?
We'll add a list of speed tips and tricks to the docs for the OR module but it may not cover your situation.
STeven.
|
|
|
|
Trop from Switzerland [6 posts] |
12 year
|
I have attached some training images i was using.[Car (1)] I am new just trying to use the Haar features to detect cars on google earth, I need to use multiple different car training images so that it becomes more robust, but then it gets realy slow
  
|
|
|
|
Trop from Switzerland [6 posts] |
12 year
|
Here is a result that I obtained for a frame
|
|
|
|
Anonymous |
12 year
|
Trop,
You may be better off using the Shape method as that would only require one template per car and not a template for any orientation of a car. The haar method will not auto-rotate the template in order to find cars at different orientations. See if switching to that method and removing the rotated templates helps to speed things up.
STeven.
|
|
This Thread
