Adi's profile - activity

We can continue the discussions here: https://github.com/opencv/opencv/issues/12974

I misunderstood the meaning of getVarCount(). I thought it was related to the number of examples in the training phase:

I misunderstood the meaning of getVarCount(). I thought it was related to the number of examples in the training phase.

Yes, StatModel is an abstract interface, but all derived classes expect a model of some known size.

See above tho this is a question about API, so the code is rather simple/obvious...

Checking the Sample Size Input for `cv::ml::StatModel::predict()` Loaded from File? I'm loading an ML trained model from

Checking the Sample Size Input for `cv::ml::StatModel::predict()` Loaded from File? I'm loading an ML trained model from

Checking the Sample Size Input for `cv::ml::StatModel::predict()` Loaded from File? I'm loading an ML trained model from

Checking the Sample Size Input for `cv::ml::StatModel::predict()` Loaded from File? I'm loading an ML trained model from

Checking the Sample Size Input for `cv::ml::StatModel::predict()` Loaded from File? I'm loading an ML trained model from

Checking the Sample Size Input for `cv::ml::StatModel::predict()` Loaded from File? I'm loading an ML trained model from

Checking the Sample Size Input for `cv::ml::StatModel::predict()` Loaded from File? I'm loading an ML trained model from

For completeness, this also works: img = img.mul(Scalar(1.2,0.5,.2));

You should past this answer here too: https://stackoverflow.com/questions/4...

@berak: Yeah, this seems to work. IIUC, this is essentially the same as using cv::transform(). It is matrix multiplication after a reshape.

It seems that given a multichannel image, e.g. BGR, img I cannot do this:

img *= cv::Scalar(1.5,0.5,2.1)

I'd like to scale each channel by a different float factor. Is there a simple way to do this? I could use cv::transform() but that seems like overkill (I also obviously don't want to iterate on all the pixels).

Any suggestions?

I cam upon the undocumented color conversion codes: CV_Luv2LBGR and CV_Lab2LBGR (and their LRGB siblings). Apparently, they should return the L channel in addition to the BGR channels.

However, I cannot seem to get them to work, I suspect the code actually ignores this and simply returns BGR.

How can I use them and why aren't they documented?

Answering myself: use cv::inRange().
See also this SO answer.

Given a multichannel image (e.g. BGR), what's the most efficient way to generate a (binary) 1C mask for pixel value equality?

cv::compare() works only for single channel images.

Using op== yields a 3C mask.
I can then reduce that but that seems like a wasteful second pass.

Is there a single-pass way of doing this?

Hi,

here's a code sample without cv::reshape() or cv::reduce() :

cv::transform(m_in, m_out, cv::Matx13f(1,1,1))

If your raw data is a normal image, then any lossless format that supports your raw data depth should be fine. Try PNG. It is a lossless image format.

Indeed ;-)

The docs tell of a function guidedFilter() in OpenCV (with some more goodies along with it).
I can't find anything about this in the 3.0.0 repo.

Is this in some special branch?

I added a sample. Both FAST and AGAST are very fast corner detectors, or center-surround contrast estimators. They're just the right approach for what I need, and in fact give pretty good results already. I am looking for generalize them just a bit more.

I want to detect tiny dark points/dots in a color image.
Is there an easy way to get FAST and/or AGAST to measure the score as the difference between the min-of-channels in the center p and the max-of-channels in the points around each center?

If this is not possible with the current API, how would one change FAST or AGAST to accept 2 images, one for the center and one for the outline?

Here's an example. It has two darker faint spots:

The background is also noisy (and even more in other cases).
FAST and AGAST can actually detect the spot closer to the center, but the one on the bottom-right corner cannot be detected without detecting a lot of noise.

Since FAST/AGAST work only on gray-scale images, it would be a great addition if we could use 2 inputs, one for the center-pixel and one for the surrounding pixels.