I am trying to use the example from here to detect a stop sign.
But when I change the image to search for stop signs in there are a huge number of false positives.
Am I doing something wrong?
$\begingroup$
$\endgroup$
2
-
$\begingroup$ Hi i'm also trying the trainCascadeObjectDetector for detecting people but it doesn't show me a good results someone could help me plz . my data set contains 270 negative images and 450 postive images . other question please i need to know haw to test haar feature in trainCascadeObjectDetector('personDetector.xml',data ,negativeFolder,'FalseAlarmRate', 0.2, 'NumCascadeStages',18); i know that hog is the default feature . thank you :) $\endgroup$– narjesMar 25, 2015 at 10:31
-
$\begingroup$ Please do not use answer slots to ask similar questions. You can use the comments block under the question, or start a new thread if your case is different enough. $\endgroup$– sansuisoMar 25, 2015 at 11:55
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
7
You are not doing anything wrong. This particular example is there for a quick illustration, and it does not produce a very good detector. It trains a 5-stage detector from a very small training set. A decent detector may have 20 stages or more, and you would need thousands of positive samples and negative images to train it.
Take a look at this tutorial for more information.
-
$\begingroup$ Okay, thanks for letting me know that. I have noticed that it actually works pretty good on images that contain stop signs about the same size (in pixels) as the test image. Do you happen to know of a more in depth explanation of how to train a cascade object detector in MATLAB than the one you provided? $\endgroup$– user8919May 29, 2014 at 4:53
-
$\begingroup$ Adding to what I previously wrote: So I guess you would say that the example is not very robust to changes in scale. I have read that the algorithm used is pretty robust to changes in scale though. Should I have to train the classifier with many different scales of the same images? $\endgroup$– user8919May 29, 2014 at 5:03
-
$\begingroup$ In this example the ObjectTrainingSize is set automatically to [33,32]. So stop signs smaller than that will not be detected, but the ones bigger than that should be. If they are not, then try decreasing ScaleFactor parameter of vision.CascadeObjectDetector. $\endgroup$– DimaMay 29, 2014 at 13:56
-
$\begingroup$ I don't know of a better tutorial off hand. Is there anything specific that is not clear to you? Please keep in mind, that this is a toy example. You need vastly more data to get a useful object detector. Also keep in mind that stop signs at oblique angles are not likely to be detected, because they would have a different aspect ratio. $\endgroup$– DimaMay 29, 2014 at 13:59
-
$\begingroup$ So the particular object I am trying to train the detector for has a pretty constant aspect ratio. But for my application the object is very wide compared to tall. So when I turn the object 90 degrees the object has the opposite aspect ratio. So do I need to train the detector with examples turned 90 degrees, or does the algorithm automatically train for orientation shifts in the plane like that? $\endgroup$– user8919May 29, 2014 at 17:40