Comments by BupeChombaDerrick (Top 75 by date)

Your initialization for some variables might be causing the problem. Check my modified code.

Check my modified code. Added this line if (newsx == gx && newsy == gy){
found = true;
}

Check my modified code. But insure that each method is doing what it is supposed to be doing. I have added a "found" flag to terminate the loop when the go state is reached. Also set max_itr = bitmap.Height*bitmap.Width, so that the algorithm can go through every pixel if possible.

Can you provide the implementations of vs?

Looks like "vs(si, sj, newsx, newsy);" should find the eight neighbors but it is being invoked 8 times in the loop, put it outside before the loop.

According to your provided pseudocode, you need to find the eight neighbors of a given point, where,in your actual code, is this happening?

Is the "bitmap" object properly allocated? it's dimensions should much "state" array.

Maybe you should take a look at this http://www.codeproject.com/Articles/42258/Particle-swarm-optimization-for-function-optimizat . Your algorithm must have some errors.

Your code risks going in an infinite loop because of this, if (state[newsx, newsy] == state[gx, gy])
iter_no = iter_no + 1;
You should increment the loop counter outside the if statement. For speed use ++i in your for loop increments

You know I'am developing a robotic navigation system and a general problem solver for intelligent machines in cases involve sensory processing, the approach I want to use is now based on perception rather than optimization. By perception, I mean, the system must base it's action on matching an observed scene to thousands of models in a database, these models will have metadata about the best navigation route and anticipate object and scene dynamics such as moving objects e.t.c. This is a general thinking process I'am developing for machine learning. I know this has nothing to do with your question, just wanted to share this.

An android app that has an online panorama album. I have already created the app but I want to update it so that users can upload there panoramas to a public online album. Only the album should be online, the app is usable with an offline panorama album. Like this one for photaf http://www.photaf.com/

My gcc version code is 4.5.3, I just installed the latest ndk version and still the file is missing. So I guess it's just patching then. Thanks again.

Thanks, are there any other libraries for Android for multiprocessor programming? I know of renderscript, but I can't use a c++ library with it, I'am trying to run image processing algorithms on multicore processors since there are so many devices with such types of processors.

Okay, then whether to use synchronization techniques or not depends on the complexity of the response function. I have given your answer +5, but if you want me to accept it then change it so that it includes some of the discussion here. The "chunked" approach may work but for a simple kernel implementation it maybe an overkill. Thanks for the taking the time to help out.

Okay so I shouldn't worry about race conditions or false sharing? I just have to choose to share the source(src) and destination(dst) bitmaps among the threads and openMP will handle the synchronizations?

I'am currently looking at http://staff.city.ac.uk/~sbbh653/publications/OpenMP_SPM.pdf they seem to be using the full image without a "chunked" approach. Only adding a "#pragma omp parallel for" should have been sufficient. After adding "shared(src,dst)" it seems to work but don't know why.

Typical values are width = 640, height = 480, thus about 300K pixels, the Kernel interface has two bitmaps as arguments the one designated as "dst" is mostly writeonly and "src" is readonly. Thus bitmap pixels for "dst" are expected to be modified by the threads, so one thing is that shared arrays of data like this are hard to parallelize but yet they are inherently parallel.

Actually the bitmap data is just a 1D array of pixels, so the data is just one big strip.How about the race conditions with the bitmaps? Especially the "dst" bitmap needs to be synchronized. I tried using critical within the implementations of "Kernel" when writing to the bitmap but yielded no benefits. For example

// Kernel for changing bitmap to gray scale
class GrayScale: public Kernel {

public: void response(Bitmap &src, int x, int y, Bitmap &dst)
{
Pixel &sp = src.pData[x + y*src.width];

Pixel &dp = dst.pData[x + y*dst.width];

unsigned char gray = static_cast<unsigned char="">(0.3*sp.red + 0.5*sp.green + 0.2*sp.blue);

#pragma omp critical
{
dp.red = dp.green = dp.blue = gray;
}
}
};

where

// Pixel data for ARGB_8888 bitmap format
struct Pixel {
unsigned char red;
unsigned char green;
unsigned char blue;
unsigned char alpha;
};

Well the question is not about the way variables are declared, but how the processing is done, Code 1 is accessing a pointer variable every time in the loop, while Code 2 is only doing it once. It turns out that both codes will run equally fast when compiled in release mode since the compiler will optimize it for me, thus to further speed things up, parallel processing is the only answer :)

Thank you. I guess I have to focus more on hardware acceleration.

Just thought I had to tune things up myself, maybe some compilers ain't that intelligent :)

I have already created my own computer vision algorithms for mobile and I have this app using my vision algorithms https://play.google.com/store/apps/details?id=com.facebook.rethinkvision.Bimostitch, I want to optimize the code on low and high level. You are right I guess I need to do parallel processing, I'am currently doing some research on that topic, any links to good literature on parallel processing can be helpful.

Does accessing the member variable within a loop affect the code performance? I'am writing computer vision code and I want to fully optimize it.

You can try heapsort, check my updated solution :)

Yes I have "import android.support.v8.renderscrpt.*" line instead of "import android.renderscript.*" but APK installation fails on android 2.2.

In String[] names = (String[]) members.toArray();

What is members? Maybe it is the cast that is the problem.

I think the improvement to the solution can now handle the case of 2 items having the same value.

You are welcome :) I have slightly improved the solution.

Still I see AlarmManager and Service in your answer :).

Use a cursorAdapter for that.

You are welcome :).

No, the emulator is exactly like a real device, so if it does not work on the emulator then it won't work on the real device. So it will return null even on the real device. Your emulator and real device might be running different versions of android OS, check the versions.

Caused by: java.lang.NullPointerException
com.example.temp.MainActivity$Asynctask.doInBackground(MainActivity.java:114) check whats on line 114 in your code, that's where the problem is.

You are welcome, your link looks like a better solution :).

It seems you have to insert the image using "Images.Media.insertImage()" for it to be saved to SD card and so that the MediaStore can create a path for it. The intent seems only to return the raw image data. But I'am not very sure.

You are welcome.

If you want very fast blurring implement box filters using integral images (summed area table) http://computersciencesource.wordpress.com/2010/09/03/computer-vision-the-integral-image/.

I must say I have learnt a lot about threads just from this answer, thanks Sergey for the inform, it might come in handy. A vote +5.

Don't know much about threading, but I think I get Sergey's point, threads just run independent of one another, so even if one thread crushed, it wouldn't affect the other running threads. And if a processor is multicore, then seperate threads can run on separate cores, but if the number of threads exceeds the number of cores then the threads become overheads rather than improving performance of the program, so in some cases threading can improve performance while in others it can hurt performance, especially if more than the necessary number of threads are running. I'am right? :-)

You are welcome. :-)

You can use openCV to access kinetic RGB-D image,where D = depth, so I was assuming that using the depth map or channel can help in segmenting out the human from the background and then you can train neural networks on specific spatial and temporal signals from the segmented training examples. I have not done much of neural networks with temporal pattern recognition but maybe these two links below can help you with the kinetic sensor and openCv and spatial and temporal pattern recognition respectively.

http://opencv.willowgarage.com/wiki/Kinect

http://www.bu.edu/asllrp/stp.pdf

Check the other link in the answer for Emgu Csharp pedestrian detection code, you still need that pdf book though :-)

The only problem is that Hough Transform might not work on distorted circles or not perfectly circular object.

You are welcome.

For starters use noiseless images so that the contours are connected per single digit and then you can group those pixels. For some details in grouping see http://stackoverflow.com/questions/2378579/opencv-grouping-white-pixels

But if you want your app to deal with images directly from a web cam then your app needs to do a lot of pre processing to identify possible locations of individual characters.

For now I recommend you start with a sliding window technique, first train a classifier and apply the classifier at all possible locations and scales using an image pyramid.

What exactly are you not getting?

C++ is better with openCV, an app developed using C++ will definitely run faster than one developed using C#, but it depends on the computational demands of your app. If you think that your app will need optimized code i suggest you use C++ but if not then you can use C# because it is simpler to develop an app with C# than C++.

Thanks :)

thumbs up

you are welcome.

ROI is Region Of Interest. Anyways who knows exactly what it means here in the question :-)

The nose,mouth and eyes are all internal features so they are easier to detect even when the face is in a complex background, the ear on the other hand is external so ear detection can be erroneous when the head is in a complex background.

Thanks RaisKazi

Maybe try to enlarge the image to a reasonable size then use the same SURF code.

I think the 10x11 is too small for an algorithm like SURF, for such small images. you can just use normalized template matching.

check out http://www.cplusplus.com/forum/lounge/17053/

wow i'am impressed. I think on the part of clicks you need to simulate mouse events. check http://msdn.microsoft.com/en-us/library/ms171548.aspx

Hey i wish i could help ... but i did look at the code. I couldn't find a reason why the program was not working. Try downloading and modifying this code http://www.codeproject.com/Articles/26897/TrackEye-Real-Time-Tracking-Of-Human-Eyes-Using-a
Just use template matching for closed eyes and set a detection threshold.

For example, you want to run a Haar Face detector but with a .NET C++/CLI GUI , create a class called FaceDetector in a File FaceDetector.h and implementation file FaceDetector.cpp.

class FaceDetector
{
public: FaceDetector();

public: ~FaceDetector();

public: void readFromBitmap(Bitmap ^bmp); //convert to cv::Mat frame

public: void copyToBitmap(Bitmap ^bmp); //copy from cv::Mat frame to bitmap

public: void detectAndDraw();

// TO-DO: write methods that use openCV haar classifier for Face detection
// get data from a .NET Bitmap using readFromBitmap() and run face
// detection using detectAndDraw() then copy results to Bitmap using
// copyToBitmap() and you can then show the results on a Picture Box
public: Mat frame;
};

then you can include a header FaceDetector.h in your .NET C++/CLI project

then just before the form class code declare a FaceDetector object.

#include "FaceDetector.h"

FaceDetector faceDetector; //global declaration

// C++/CLI code here

then whithin the Form class you can easily call reference the faceDetector object.

or you can use Emgu CV .NET wrapper for openCV see http://sourceforge.net/projects/emgucv/

You have to train your own haar classifier for that, but you can use template matching... or train an SVM (Support Vector Machine) classifier on open and closed eyes for both right and left eye ... detect faces using haar classifier and then run the SVM open and closed eye detector on those regions with faces. There is no easy way around this and so you need to put in the effort to train your own classifier.

i think "haarcascade_eye.xml" is corrupted try "haarcascade_eye_tree_eyeglasses.xml" for eyes and "haarcascade_frontalface_alt.xml" for faces. see the code below for face detection,its a modification of your code. I ran it using OpenCV 2.3.1,though i did not include mouse(int delta_x,int delta_y) function here,i believe that's for you to do. But i see you want to detect eyes right away... though that will be error prone... do face detection using fast haar detector and then eye detection probably using template matching because haar classifiers don't seem to perform better than template matching on eyes. And use cvSize(20,20) for better results but you can always check the size in the xml file itself so don't use any size.

#include "opencv2/opencv.hpp"

using namespace cv;

CvHaarClassifierCascade *cascade;

CvMemStorage *storage;

void detectFaces( IplImage *img );

int main( )
{
CvCapture *capture; IplImage *frame = 0;
int key;
char *filename = "haarcascade_frontalface_alt.xml";

capture = cvCreateCameraCapture(0);

storage = cvCreateMemStorage(0);

cascade = (CvHaarClassifierCascade*)cvLoad( filename,storage, 0, 0 );

assert( cascade && storage && capture );
cvNamedWindow( "video", 1 );
while( capture )
{
frame = cvQueryFrame( capture );
if( !frame )
{
fprintf( stderr, "Cannot query frame!\n" );
break;
}

detectFaces( frame ); key = cvWaitKey( 25 );
}

cvReleaseCapture( &capture );
cvDestroyWindow( "video" );

cvReleaseHaarClassifierCascade( &cascade );
cvReleaseMemStorage( &storage );
return 0;
}

void detectFaces( IplImage *img )
{
int i;
CvSeq *faces = cvHaarDetectObjects(img,cascade,storage,1.15,3,0,cvSize( 20, 20 ) );
for( i = 0 ; i < ( faces ? faces->total : 0 ) ; i++ )
{
CvRect *r = ( CvRect* )cvGetSeqElem( faces, i );
cvRectangle( img,cvPoint( r->x, r->y ),cvPoint( r->x + r->width, r->y + r->height ),CV_RGB( 255, 0, 0 ), 1, 8, 0 );
}

cvClearMemStorage( storage );
cvShowImage( "video", img );
}

please state clearly what you mean by "not working properly" ... because if the code is error free then the algorithm must not be okay. I'am taking a look at the code will get back to you after i do some analysis!

Totally agree with Philippe, eye blinks are too fast, besides what use can you make of such a system (enlighten me), try using haarcascade_eye.xml file that comes with openCV to detect eye positions as a pre-processing stage and then see if you can invent a classifier to tell whether a person blinked. You can only capture such high speed events (blinks) using expensive high speed cameras which makes your project unfeasible and i doubt if the system you which to design can process frames at such high frame rates to be applicable.

Thanks for the time you've taken to help.I really appreciate it.

Similar to Google Goggles get a snap shot or any digital image and upload the image and press enter sit back wait for the results. Has of now the system is using a web-cam in real-time to recognize images or objects presented to it, i want it to be hosted on a server so that i can use it as a visual search engine. I want to develop a Visual encyclopedia so that users can get info about visual objects directly using this system.

A vision system is an image processing software, in this case i have developed an image indexing program that can help index internet images and make them searchable. The system uses high-level features for this purpose unlike the existing ones which use intermediate or low-level features. I have tried the system on a small database but now i want to see how it can perform indexing a few thousands or millions of images existing on the internet by letting users interact with the system via a web browser.

Thanks.

Anyway of communicating the finds of the research but protecting the technology from being used commercially by someone without permission would suffice.