Optimal size (dimensions) of negative learning images for HaarTraining?

classic Classic list List threaded Threaded
2 messages Options
Reply | Threaded
Open this post in threaded view
|

Optimal size (dimensions) of negative learning images for HaarTraining?

beretux
Hello,
I want to learn OpenCV to detect some simple objects (for example
http://tinyurl.com/bwolhf) on scanned
paper. I have many small positive learning images - 38x38 pixels (100
scanned and cropped + every image distorted a bit by 'createsamples'
app 50 times --> 100*50 positive images).

I'm interested, what is good for negative learning images - is better
to have:
* ~100 of images with high-resolution (~1700x2300px) (example)
or
* ~10000 of images with low-resolution (for example 38x38px - same
size as positive images) automatically and randomly cropped from
high-resolution images
(example http://tinyurl.com/c8qoyh)

or something between these extremes? somebody knows, what is better
(and why)?

Thank you in prediction,
Petr Kotek

Reply | Threaded
Open this post in threaded view
|

Re: Optimal size (dimensions) of negative learning images for HaarTraining?

wl2776
Administrator
> I'm interested, what is good for negative learning images - is better
> to have:
> * ~100 of images with high-resolution (~1700x2300px) (example)
> or
> * ~10000 of images with low-resolution (for example 38x38px - same
> size as positive images) automatically and randomly cropped from
> high-resolution images
> (example http://tinyurl.com/c8qoyh)
>
> or something between these extremes? somebody knows, what is better
> (and why)?
>

I vote for the second variant.
Your samples form 1444-dimensional space (38*38 = 1444). They form a cloud of points
in that space, and the boosting tries to build several hyperplanes, forming polyhedron,
surrounding these points. More negative samples - the cloud boundary is more accurate.