kaipyroami's profile - activity

I am using iOS (iPad) I am at the current version.

That is a good idea, I am just trying to limit adding any more extra processing if I can help it. I was thinking that the camera input might be easier to correct than flipping the mat.

If you take the first image you posted and rotate it 180 degrees that is what I am getting. 0,0 should be in the top left but is in fact in the bottom right.

Fixed while overriding CvVideoCamera. Not entirely sure why it is working now though. :/

Are you only able to detect open eyes? But you can detect the left and right eye only when open?

Are you only able to detect open eyes? But you can detect the left and right eye only when open?

Just an observation, though it might solve the problem. You are using an extremely small nimber of sample images. I am using 2,000+ negatives and 3,000+ positives. Perhaps there are not enough images to start one full stage?

You need to override - (void)updateOrientation and - (void)layoutPreviewLayer.

CvVideoCameraMod.h

 #import <opencv2/highgui/cap_ios.h>

@protocol CvVideoCameraDelegateMod <CvVideoCameraDelegate>
@end

@interface CvVideoCameraMod : CvVideoCamera

- (void)updateOrientation;
- (void)layoutPreviewLayer;

@property (nonatomic, retain) CALayer *customPreviewLayer;

@end

CvVideoCameraMod.m

#import "CvVideoCameraMod.h"

 #define DEGREES_RADIANS(angle) ((angle) / 180.0 * M_PI)

@implementation CvVideoCameraMod

- (void)updateOrientation;
{
    NSLog(@"Would be rotating now... but I stopped it! :)");
    self.customPreviewLayer.bounds = CGRectMake(0, 0, self.parentView.frame.size.width, self.parentView.frame.size.height);
    [self layoutPreviewLayer];
}



- (void)layoutPreviewLayer;
{
    if (self.parentView != nil)
    {
        CALayer* layer = self.customPreviewLayer;
        CGRect bounds = self.customPreviewLayer.bounds;
        int rotation_angle = 0;

        switch (defaultAVCaptureVideoOrientation) {
            case AVCaptureVideoOrientationLandscapeRight:
                rotation_angle = 180;
                break;
            case AVCaptureVideoOrientationPortraitUpsideDown:
                rotation_angle = 270;
                break;
            case AVCaptureVideoOrientationPortrait:
                rotation_angle = 90;
            case AVCaptureVideoOrientationLandscapeLeft:
                break;
            default:
                break;
        }

        layer.position = CGPointMake(self.parentView.frame.size.width/2., self.parentView.frame.size.height/2.);
        layer.affineTransform = CGAffineTransformMakeRotation( DEGREES_RADIANS(rotation_angle) );
        layer.bounds = bounds;
    }
}
@end

This works. By using "pathForResource" of NSBundle.

NSString * const TFCascadeFilename = @"haarcascade_TF";//Strings of haar file names

NSString *TF_cascade_name = [[NSBundle mainBundle] pathForResource:TFCascadeFilename ofType:@"xml"];

 //Load haar file, return error message if fail
 if (!TF_cascade.load( [TF_cascade_name UTF8String] )) {
     NSLog(@"Could not load TF cascade!");
 }

I am having trouble with the input from my camera being rotated 180 degrees. In other words, the image is upside down when my cvPoints and text are correct. But when the image is right side up the X and Y locations are referenced off of the bottom right corner instead of the top left. And text is upside down.

Aside from "AVCaptureVideoOrientation" which rotates everything together is there any way to rotate the camera input relative to OpenCV?

Is there anyway to stop openCV from rearranging camera input to align with the devices current orientation? I have an app that is set to a single orientation only, but as you move it around the openCV image keeps re aligning to match "right side up".

Has anyone had any trouble using object detection and iOS? This code works on eclipse with OSX and a webcam. I am able to get the "detection box" drawn on the image so the image is being passed properly to the processImage and detectAndDisplay methods. By adding break points I am able to see that the code appears to execute as expected. I am however not able to get an object that is detected on the webcam to be detected on the iPad.

I am thinking that since the image is passed by reference it might be getting written over before the detection is completed, but I am not sure.

In ViewDidLoad I load the haar files:

  TF_cascade_name = "haarcascade_TF.xml";
    FF_cascade_name = "haarcascade_FF.xml";

    TF_cascade.load( TF_cascade_name );
    FF_cascade.load( FF_cascade_name );

Then here is my image recognition code:

    - (void)processImage:(Mat&)image
{  
    offset_x = 400;
    offset_y = 150;

    cv::Rect myROI(offset_x, offset_y, 500, 200);

    cv::Mat croppedImage = image(myROI);
    [self detectAndDisplay:image detectionROI:croppedImage];     
}

/** detectAndDisplay */
- (void)detectAndDisplay:(Mat&) frame
             detectionROI:(Mat&) ROI_frame
{
    std::vector<cv::Rect> FFs;
    std::vector<cv::Rect> TFs;

    cv::Mat frame_gray;

    cvtColor( ROI_frame, frame_gray, CV_BGR2GRAY );//convert to grayscale
    equalizeHist( frame_gray, frame_gray );

    //-- Make Detection Box
    rectangle(frame, cvPoint(offset_x,offset_y), cvPoint(offset_x + 500,offset_y + 200), cvScalar(0,255,0),2, 1);


    //-- Detect syringe

    TF_cascade.detectMultiScale( frame_gray, TFs, 1.1, 2, 0|CV_HAAR_SCALE_IMAGE, cv::Size(30, 30) );

    for( int i = 0; i < TFs.size(); i++ )
    {
        cv::Point center( (TFs[i].x + TFs[i].width*0.5) + offset_x, (TFs[i].y + TFs[i].height*0.5) + offset_y );
        ellipse( frame, center, cv::Size( TFs[i].width*0.5, TFs[i].height*0.5), 0, 0, 360, Scalar( 255, 0, 0 ), 4, 8, 0 );

        Mat faceROI = frame_gray( TFs[i] );
        std::vector<cv::Rect> eyes;        
    }


    FF_cascade.detectMultiScale( frame_gray, FFs, 1.1, 2, 0|CV_HAAR_SCALE_IMAGE, cv::Size(30, 30) );

    for( int i = 0; i < FFs.size(); i++ )
    {
        cv::Point center( (FFs[i].x + FFs[i].width*0.5) + offset_x, (FFs[i].y + FFs[i].height*0.5) + offset_y );
        cv::ellipse( frame, center, cv::Size( FFs[i].width*0.5, FFs[i].height*0.5), 0, 0, 360, Scalar( 0, 0, 255 ), 4, 8, 0 );

        Mat faceROI = frame_gray( FFs[i] );
        std::vector<cv::Rect> eyes;
    }
}

Anyone know if it has to do with "lockForConfiguration"? I am not able to implement it in a CvVideoCamera object, however many other threads show this being a solution to a non-responsive property.

http://docs.opencv.org/trunk/doc/tutorials/ios/video_processing/video_processing.html#opencviosvideoprocessing From the above tutorial are there any known issues with the following line of code and iOS6?

self.videoCamera.defaultAVCaptureVideoOrientation = AVCaptureVideoOrientationPortrait;

The orientation seems to be unchanged and is always in portrait even when I use AVCaptureVideoOrientationLandscapeRight.

Using the tutorial posted here.

On this line:

self.videoCamera.delegate = self;

I am getting the error:

Assigning to 'id<cvvideocameradelegate>' from incompatible type 'Camera_FilterViewController *const __strong'

Any pointers on what to check out first?