image questions in regards to training cascade classifier

asked 2019-03-20 01:07:39 -0600

wgb-2019
21 ●3

So far I have created two LBP classifiers that do a decent job of identifying the shapes I need identified however, I would like to increase the accuracy and I have the following questions:

Should all the positive and negative images be of the same size? i.e. 64 x 64 or 32 x 64 etc...
If #1 is a no, do all the positive images or all the negative images have to be the same size or ratio?
Should the positive image be cropped to only include what I am looking for or should it have some background ?
Should the negative image be totally different (ex: out of context) from the positive or should it be similar (ex: display the background but not the positive image)
Are there any rules of thumb on how many images I should use. So far I have used 1000 positive and 1000 negative.
Does image orientation play any role i.e. left to right vs right to left?

Thanks in advance WGB

answered 2019-03-20 01:48:30 -0600

berak
32993 ●7 ●81 ●312

no
no, again. you have to give the desired positive size to the createsamples tool, and make an annotation txt file with image / object rects (to be cropped later), e.g. using the annotation tool fom the apps. again, both positive and negative images can have arbitrary size
no, the createsamples tool will do that for you
you need a realistic background (it has too learn that, too !)
the more, the better. (don't try with 50 only)
yes. you can vary the pose, say, +- 20%, but if you need a detection from the right, and also a detection from the left, you need to traiin 2 seperate classifiers

have a look at the docs, it's all explained there !

last: it's 2019 now, and you probably should retrain a (SSD style) cnn for your purpose. the cascades are somewhat "on the way out" , e.g. latest opencv does no more contain the tools to train those.

edit flag offensive delete link

Comments

The reason I choose LBP is due to its speed. I need to scan around 450 frames (300 fps) and extract the results almost in real time. How fast are CNN's?

wgb-2019 ( 2019-03-20 02:20:43 -0600 )edit

yea, good luck. just be aware of it's limitations, and, that code for this might not be maintained in the future.

berak ( 2019-03-20 02:28:23 -0600 )edit

add a comment

image questions in regards to training cascade classifier

1 answer

Comments

Links

Question Tools

Stats

Related questions

image questions in regards to training cascade classifier edit

1 answer

Comments

Links

Question Tools

Stats

Related questions

image questions in regards to training cascade classifier