Ask Your Question
0

Opencv training, difference between one image and set of image.

asked 2016-01-31 06:18:21 -0600

Corto gravatar image

Hi all, I have create a road signal classifier using four signal (Stop, turn left, turn right and roundabout), i create the .vec file using one image for every road signal (one for stop, one for left ecc.) with opencv_createsamples, and with the file i training the classifier. I use the .xml file in one app for detect and track the signal with android but i have a problem, the classifier do 14 (at last) stage and never reach 20, and when i try to detect a photo of one (i have take the photo around) of this the classifier don't detect it. The question is: Is it better create a classifier with one image (With createsamples) or do a lot of a photo object? why the classifier does not work with real photos? Thanks for help!

edit retag flag offensive close merge delete

Comments

i have never seen a good classifier, that was trained on a single image only. rather get a few hundred images.

then, i doubt, that doing 5 cascade detections is a good idea at all. since a single detection on android already takes almost a second, this approach won't scale.

do your signs have the same shape ?

berak gravatar imageberak ( 2016-01-31 09:53:11 -0600 )edit

The stop signal is hexagonal and the other are circle. With stop sign, turn left and turn right work good, they give me a little problem but work, only with roundabout i have problem. I think the error is in the images used for training the classifier. How i can get a few hundred image? With createsamples it's possibile?

Corto gravatar imageCorto ( 2016-02-01 11:14:31 -0600 )edit

By grabbing a camera, get on the road and collect sign images in real life conditions ... this is about the only thing that will work decently!

StevenPuttemans gravatar imageStevenPuttemans ( 2016-02-03 02:00:06 -0600 )edit

i think i do like this, how many image do you think i take? he must have the same size or don't care? and in the image there is only the signal?

Corto gravatar imageCorto ( 2016-02-04 03:41:16 -0600 )edit

You grab like 500 images, then you use the opencv annotation tool to select the regions of interest, who will in turn get resized to a default width and height using the opencv create samples tool.

StevenPuttemans gravatar imageStevenPuttemans ( 2016-02-04 03:57:58 -0600 )edit

ok, i take 500 photo of image, for the tool i use Haartraining_Stuff. I think is the same! No?

Corto gravatar imageCorto ( 2016-02-05 09:50:59 -0600 )edit

1 answer

Sort by ยป oldest newest most voted
0

answered 2016-02-14 04:35:46 -0600

Corto gravatar image

I resolve this, i use 150 positive image and 200 negative image and it work's pretty well.

edit flag offensive delete link more

Question Tools

1 follower

Stats

Asked: 2016-01-31 06:18:21 -0600

Seen: 332 times

Last updated: Feb 14 '16