GrumbleLion's profile - activity

why would main call pthread_exit ? Also does the code pass a reference to part of the image or actually copy the image data to the vector?

FYI: you can also use cv::BORDER_REPLICATE to make the corners white.

Wrong in what way? Is your template color balanced? what method? is it normalized?

Do you have code to check to see that both the template and the image are the same format? Also the heat map has to be created just right.

%1 was used as a file reference name in MSDos batch files. I am not sure what it would have in the python import file.

I am using the Region of Interest to copy one smaller image on to another.

cv::Rect roi( cv::Point( 0, 0 ), cv::Size( bla.cols, bla.rows ));

cv::Mat destinationROI = MyImageBuffer( roi ); // make a pointer to this area.

bla.copyTo( destinationROI ); // draw to that pointer

I've had a few thread glitches while using OpenCV.

The library itself is thread safe in that you can have multiple calls into the library at the same time, however the data is not always thread safe.

This means that if you have to pass a mat image between threads you have to use a mutex and call .CopyTo() to force a copy instead of a reference.

Note: Std::Vector is also not thread safe under all conditions as well.

I don't know enough to give you all the pros/cons between those two. There are many new articles about newer detectors, however it all depends on what you are trying to detect and what the situation is. In my case the camera is in a fixed place and the target image is provided so I went with match template and used a background scan to try different sizes and rotations. I then added a sub pixel routine to increase the accuracy.

Given that fish can be at odd angles I would think looking for color would be a better place to start.

You could calibrate the camera and apply correction.

You could convert the rect into two triangles and then check to see if the point is inside the triangles. http://www.blackpawn.com/texts/pointinpoly/

NTSC scans odd and then even lines. You could take two frames and average them.

First you have to create the 2x3 rotation matrix.

cv::Point2f center(((float)imageSize.width * 0.5f), ((float)imageSize.height * 0.5f)); cv::mat RotMap = cv::getRotationMatrix2D( center, Rot, 1.0);

then rotate your image

cv::warpAffine(ImageMatIn, RotatedImageMatOut, RotMap, CameraAvgImage.size(), cv::INTER_LINEAR, cv::BORDER_REPLICATE , cv::Scalar(127,127,127));

I am just playing with this now trying to make a better corner detector and I am guessing you first have to increase the BlockSize, however I don't see a parameter for that.

I don't see a cap.open

you could typecast mat into an array.

Your second image contains data that does not exist in the first one. This means that either you are displaying the first one as a BW image or your second one is the first one.

ok... then to get the rotation you can create a weighted filter. or for a crude value you could find the points that are farthest from the center.

I don't see it being able to detect any arbitrary whole number, but it can detect each number and then your code would have to string them together.

See The HighGui examples of how to open a camera.

Does the image change shape or size? The first step is to use one of the object detectors. If you are generating the shape and it's BW you could use an average of the X and Y of all the black pixels to find it's center and then take the average of x and y for all the black pixels that are in the 4 quadrants from the center to get the rotation.

why are you using "true color" conversions that then depend on your monitor settings?

The Asus Xtion WebCam gives you a point cloud so you might not even have to use OpenCV.

I ended up with a separate thread and a dual buffer system, but since my target is an I7 PC it's got plenty of time and ram for multiple image copies. In a timer based single thread system you need to find a non blocking way to check to see if a new frame has came in. If you do use cv:mat make sure you use Clone or copyTo so that the actual buffer is copied to prevent thread problems.

first you have to calibrate the spherical distortion of the camera. this only has to be done once. See the calib3D examples. Then you would use object recognition to find the triangles. from there you have to use some math and find the camera's location.

you would also have to measure the radial distortion of the camera that was used.

sounds like a SVD function.

if what the camera is looking at has targets and you know where the targets are located it is possible to work backwards and figure out where the camera is located.

The crack is not a perfect line but it does have some sections that are less noisy. I would think you would have to start with a Gaussian blur function. Then the hough functions start by measuring local line angles and then tries to stitch them together into a longer line. The noisy bits won't pass this second part of the process.

I hate to dig up this old question, however he might still have email notification turned on.

check out the new topic. http://answers.opencv.org/question/29665/getting-subpixel-with-matchtemplate/

I have a link to my source code.

that's up to you. I just commented out the code for the XML file and set the settings the way I wanted to test and then I started on my own code.

I am working on a project that uses matchTemplate and I need sub pixel accuracy and I figure I could share the resulting code.

I've found others asking the same question, but the direction I am going is different since a general curve fit will not work since the values past the top few are not stable. So the plan is to scan out and find the smallest drop and then... edit: I am thinking that a parabola fit will work better.

Another problem is that if the template has a lot of fine detail the result will also not be stable. There are two options; option one: upsize both the template and the source image by two before match template. option two: shift the template by .5 pixel and weave those results into every other row/col of the heat map.

given the different methods I am wondering if something like this would be good.

minMaxLocSubPix(

in Mat = heat map. also gives function size of source.

in point = x,y point from minMaxLoc or any other point you want to explore

    in Polarity = Let this function know if we want large or small values

in Method = what algorithm to use since mine won't be the best for every application.

out Float x,y ) returns float error value if math result is limited to prevent moving more than one pixel.

See also this post I found http://answers.opencv.org/question/5584/find-max-with-sub-accuracy/

Here is my current code. Edit: 4/28/2014 I found a bug and updated it. http://www.longrange.net/Temp/CV_SubPix.h http://www.longrange.net/Temp/CV_SubPix.cpp

In summary: My question is... how do I add it to the library to share with others?

Can't you filter it out as you read in the data and load the mat?

I wonder if a hough circle function could be modified to allow for variations of arc during the line.

first you have good lighting. Metal tends to be reflective so it's good that you have the camera above the number and the light source from the side. Next you need a image of each number. I would then think you could use one of the object detectors.

Can you control the image on the LCD? If so you could have it display the camera calibration patterns.