OpenCV for iOS

CVPR 2012

Getting your CV algorithm up and

running on iOS

Its fun
Its trendy
It can make you rich
Its impressive way to
show some technology

Plan for today

Intro; iOS Simulator vs iOS Device.

Objective-C
Make iPhone say Hello!
Make OpenCV say
(on iPhone)
Processing photos you shoot with iPhone.
Processing live video
Performance tips & tricks
References

iOS Anatomy
Apps
Cocoa Touch (UIKit etc.)
Media (CoreImage, CoreGraphics, OpenGL ES, AVFoundation)
Core Services (Databases, iCloud, GDC, )
Core OS (Unix, POSIX, )

iOS Development
+

iOS Simulator:

iOS Device:

+ $99

Device should be provisioned a.k.a. registered to be used with the particular

developer certificate

Environment setup tutorials at [Link].

iOS Simulator

iOS Device

Cost

Free*

$99 per year

Prerequisities

OSX 10.7 + Xcode

4.3.x or later

+ iOS-5 capable
device

Target CPU

x86 (SSE, no Neon)

ARM (Neon, no SSE)

Speed

fast

slower

RAM

lots of

limited

OpenGL

Photo Library, Still

Photo processing

Camera

Accelerometer

Debugging, Logging

Performance
evaluation

Getting started

Register at [Link] (free)

Mac + Xcode is enough to start: download
myriads of samples from [Link]

Enroll yourself to iOS development program

(annual fee)

Create your certificate, put to your computer(s)

Provision your device(s)
Google for and read this

iOS Language: Objective C

Objective C = Smalltalk + C
o Apples take: Objective C++ (can use classes, STL etc.)
Since iOS 5 includes ARC (automatic reference

counting)
Dynamic methods are called by name, not by
indices in vtable
Has very rich class library with dedicated root
object NSObject
.h extension for headers, .m & .mm extension for
implementation

Hello world in Objective-C

#import <Foundation/Foundation.h>!
int main(int argc, char *argv[])!
{!
//printf("\n\nHello, world!\n");!
@autoreleasepool {!
NSLog([[NSString stringWithUTF8String:"\n\nHello"]!
stringByAppendingString:@", world!"]);!
return 0; }!
}!

Its a superset of C
#import == #include with automatic guards.
[recipient messageWithArg1:A andArg2:B];
Since OSX 10.7 & iOS 5 memory is automatically managed

Hello world in iOS (1/16)

Create new Single View Application

Hello world in iOS (2/16)

Open main.m, comment off original code and put in the
hello world code shown before. Run

Xcode 4: iOS IDE

Target (simulator/device)

Display/Hide Areas

Click on the project to

change its properties, add
frameworks
Utilities Area

Frameworks
used
Navigator
Area

Text & Resource Editing Area

Debug Area

Object
Library

Cocoa(Touch) key concept: MVC

Model-View-Controller
o For simple demos Model is a part of
Controller class.

View is created visually as a Storyboard

element
For each view component (image view, button,
slider ) we add corresponding IBOutletmarked property in the controller class
For each action we add IBAction-marked
method in the controller

Hello world in iOS (3/16)

Restore main.m. Open the main storyboard, add image view

Hello world in iOS (4/16)

Open Assistant Editor (in Xcode View menu). Open ViewController.h, add
imageView Outlet property, connect it with the view in StoryBoard.

Class declaration in Objective-C

#import <UIKit/UIKit.h>!
@interface ViewController : UIViewController {!
UIImage* image;!
}!
@property (weak, nonatomic) IBOutlet UIImageView *imageView;!
@property (weak, nonatomic) IBOutlet UIBarButtonItem *loadButton;!
-(IBAction)effectButtonPressed:(id)sender;!
@end!

@interface @end delimits class declaration

there must be one and only one base class (UIViewController in our sample)
data members are declared inside {}.
@property declares a property, IBOutlet qualifier denotes outlets.
Instance methods are declared as
(<returntype>) <methodnamepart1>: (<arg1type>)arg1 [<methodnamepath2>:];

No need to declare overloaded methods, only the newly added

Actions have IBAction return type

Hello world in iOS (5/16)

Create image (e.g. white rect with text Hello, world!), add it to the project

Hello world in iOS (6/16)

Open [Link]; synthesize imageView property; add code to
load and display image on startup.

Class implementation in Objective-C

#import ViewController.h!
@implementation ViewController!
@synthesize imageView;!
- (void)viewDidLoad {!
[super viewDidLoad]; // call the superclass method!
NSString* filename = [[NSBundle mainBundle]
!pathForResource:@"helloworld" ofType:@"png"];!
UIImage *image = [UIImage imageWithContentsOfFile:filename];!
if( image != nil ) [Link] = image;!
}!
!
@end!

@implementation @end delimits class implementation

@synthesize automatically generates correct code for reading and writing properties.
- viewDidLoad is the place to put additional initialization. Xcode generates stub for it.
Some methods, including viewDidLoad, require to call the original superclass method.
Local variables are automatically destructed if they are not needed anymore
To show the image, just assign it to the image property of UIImageViewer.

Finally, run it

Lets now add OpenCV!

OpenCV for iOS

Started with GSoC 2011 project. Sort of alpha

stage now.

Thanks to C++ Objective C interoperability

native OpenCV for iOS just works.

[Link] for simpler integration

No ARM/iOS-specific acceleration yet L (planned
for OpenCV 2.5)

No camera/video support within Highgui use

iOS facilities.

OpenCV+iOS Anatomy
Apps
Cocoa Touch (UIKit etc.)
Media (CoreImage, CoreGraphics, OpenGL ES, AVFoundation)
Core Services (Databases, iCloud, GDC, )
Core OS (Unix, POSIX, )

OpenCV on iOS is just another framework

Hello world in iOS (7/16)

1.

Build framework from OpenCV source:

[Link]

2.

Add framework to your project

3.

Change project language to Objective C++.

4.

Add #import <opencv2/[Link]> to the very beginning of *.pch file.

Converting images: UIImagecv::Mat

static UIImage* MatToUIImage(const cv::Mat& m) {!
CV_Assert([Link]() == CV_8U);!
NSData *data = [NSData dataWithBytes:[Link] length:[Link]()*[Link]()];!
CGColorSpaceRef colorSpace = [Link]() == 1 ?!
CGColorSpaceCreateDeviceGray() : CGColorSpaceCreateDeviceRGB();!
CGDataProviderRef provider = CGDataProviderCreateWithCFData((__bridge CFDataRef)data);!
// Creating CGImage from cv::Mat!
CGImageRef imageRef = CGImageCreate([Link], [Link], m.elemSize1()*8, [Link]()*8,!
[Link][0], colorSpace, kCGImageAlphaNoneSkipLast|kCGBitmapByteOrderDefault,!
provider, NULL, false, kCGRenderingIntentDefault);!
UIImage *finalImage = [UIImage imageWithCGImage:imageRef];!
CGImageRelease(imageRef); CGDataProviderRelease(provider);!
CGColorSpaceRelease(colorSpace); return finalImage;!
}!
!
static void UIImageToMat(const UIImage* image, cv::Mat& m) {!
CGColorSpaceRef colorSpace = CGImageGetColorSpace([Link]);!
CGFloat cols = [Link], rows = [Link];!
[Link](rows, cols, CV_8UC4); // 8 bits per component, 4 channels!
CGContextRef contextRef = CGBitmapContextCreate([Link], [Link], [Link], 8,!
[Link][0], colorSpace, kCGImageAlphaNoneSkipLast | kCGBitmapByteOrderDefault);!
CGContextDrawImage(contextRef, CGRectMake(0, 0, cols, rows), [Link]);!
CGContextRelease(contextRef); CGColorSpaceRelease(colorSpace);!
}!

These are not super-fast functions; minimize such conversions

Hello world in iOS (8/16)

Open [Link]; insert the conversion functions after import
statement, modify viewDidLoad as shown

!
if( image != nil ) {!
cv::Mat m, gray;!
UIImageToMat(image, m);!
cv::cvtColor(m, gray, CV_RGBA2GRAY);!
cv::GaussianBlur(gray, gray, cv::Size(5, 5), 1.2, 1.2);!
cv::Canny(gray, gray, 0, 50);!
m = cv::Scalar::all(255);!
[Link](cv::Scalar(0, 128, 255, 255), gray);!
[Link] = UIViewContentModeScaleAspectFit;!
[Link] = MatToUIImage(m);!
}!
}!

Capturing/Loading Still Images

Processing fixed image from app resources is
not practical. How to load an arbitrary image?
1. Modify StoryBoard, add toolbar & load button. Add
action for button tapped event.
2. Inside the callback open the photo library and let
user select photo from there.
3. process the loaded photo and show the result.

Hello world in iOS (9/16)

Open StoryBoard, add toolbar (it will already contain a button), rename it, add the outlet
and action for the button, connect them to button.

Hello world in iOS (10/16)

Open [Link]; add the following code for loadButtonPressed
action (did you forget to add another @synthesize for button outlet?)
!

- (IBAction)loadButtonPressed:(id)sender {!
UIImagePickerController* picker = [[UIImagePickerController alloc] init];!
[Link] = self;!
if (![UIImagePickerController isSourceTypeAvailable:!
UIImagePickerControllerSourceTypePhotoLibrary])!
return;!
[Link] = UIImagePickerControllerSourceTypePhotoLibrary;!
[self presentModalViewController:picker animated:YES];!
}!

If you try to build it, you will compile error.

[Link] = self; is Cocoa way of customizing behavior of standard
classes. ViewController needs to conform to certain interfaces.

Hello world in iOS (11/16)

Fix ViewController interface
!

@interface ViewController : UIViewController!

<UIImagePickerControllerDelegate, UINavigationControllerDelegate>!
{!
UIImage* image;!
}!

Angle brackets are used in Objective-C to specify supported protocols

(a.k.a. interfaces in Java)
Instead of creating a class derived from UIImagePickerController, we
set a delegate class, but the delegate must conform to the 2 protocols
specified above. See [Link]
for details
We need to add 2 methods to implement the protocols

Hello world in iOS (12/16)

Fix ViewController implementation
- (void)imagePickerController:(UIImagePickerController *)picker!
didFinishPickingMediaWithInfo:(NSDictionary *)info {!
[picker dismissModalViewControllerAnimated:YES];!
UIImage *temp = [info objectForKey:@"UIImagePickerControllerOriginalImage"]; !
[Link] = processWithOpenCV(temp); }!
!
- (void)imagePickerControllerDidCancel:(UIImagePickerController *)picker {!
[picker dismissModalViewControllerAnimated:YES]; }!

As in the case of overridden methods, there is no need to declare the

protocol methods just implement them.
processWithOpenCV() is our function in which we copied the
processing part of didLoad method.
The delegate can be any class, not necessarily the ViewController. So
its possible to create black-box component for faster prototyping.

Lets run it!

Tip: If you run it for the first time, the photo library will be empty. You can
drag image to the simulator, which will display it in Safari. Press mouse/
trackpad for 1-2sec until the menu arrives. Save the image to library.

Adding Camera input

Camera is not supported by Simulator L

Setting up camera, retrieving frames, displaying
them, handling rotations etc. takes a lot of code.

Well re-use component created by our GSoC

student Eduard: VideoCameraController.

Its also built using delegate pattern. The

delegate will be our good old ViewController.

Camera grabs video in YUV[Link] format. Well

only process and display luminance plane.

Hello world in iOS (13/16)

Copy VideoCameraController.h/.m to the project directory and add them
to project. Add frameworks QuartzCore, CoreImage, CoreMedia,
CoreVideo.

Hello world in iOS (14/16)

Update ViewController interface part
#import <UIKit/UIKit.h>!
#import "VideoCameraController.h"!
!
@interface ViewController : UIViewController <UIImagePickerControllerDelegate,!
UINavigationControllerDelegate, VideoCameraControllerDelegate>!
{!
UIImage* image;!
VideoCameraController* videoCamera;!
}!
!
!

#import can be used with your own files and save you from writing extra
#ifndef #define #endif.
Protocols and delegates are not exclusive things in standard Cocoa
components; well-designed components may introduce custom protocols
and use delegates. See VideoCameraController.h/.m for details.
ARC works with pointers to your own classes as well (e.g. with
VideoCamera*)

Hello world in iOS (15/16)

Setup camera in ViewController didLoad method
- (void)viewDidLoad!
{!
[super viewDidLoad];!
videoCamera = [[VideoCameraController alloc] init];!
[Link] = AVCaptureSessionPreset352x288;!
[Link] = 15;!
[Link] = self;!
[videoCamera start];!
}!

VideoCamera class is super-easy to use. Leave the default resolution,

fps etc. or customize them as needed.
Using lowest possible resolution and reasonable framerate can save a
lot of power and make apps more responsive.

Hello world in iOS (16/16)

Implement VideoCameraControllerDelegate protocol
- (void)processImage:(const vImage_Buffer)imagebuf withRenderContext:
(CGContextRef)contextOverlay {!
cv::Mat gray((int)[Link], (int)[Link],!
CV_8U, [Link], [Link]);!
cv::GaussianBlur(gray, gray, cv::Size(5, 5), 1.2, 1.2);!
cv::Canny(gray, gray, 0, 30);!
}!
- (void)videoCameraViewController:
(VideoCameraController*)videoCameraViewController capturedImage:(UIImage
*)result { [[Link] setImage:result];}!
!
- (void)videoCameraViewControllerDone:
(VideoCameraController*)videoCameraViewController {}!
- (BOOL)allowMultipleImages { return YES; }!
- (BOOL)allowPreviewLayer { return NO; }!
- (UIView*)getPreviewView { return imageView; }!

processImage:withRenderContext: should process image in-place.

videoCameraViewController:capturedImage shows the result
getPreviewView should return UIImageView where the live view is shown

The final app result

Tip: To make a screenshot of an app, press home button (at the bottom
center) and without releasing it press exit button (right side of the top edg)
for a short time.

Bonus track: Face Detection

Look at FaceDetectSimple app on the USB key and try to add the same
functionality to our HelloWorld sample.
Hints:

Add lbpcascade_frontalface.xml as a resource and load it in didLoad.

You may need to transpose input image before running CascadeClassifier,

because currently orientation is not handled properly. Then transpose the
found face coordinates back to visualize results correctly.

General iOS performance tips

Do expensive initialization once (e.g. in

NSViewControllers didLoad method)

Setup your camera preview resolution and FPS

to the lowest possible values

Use [Link]
Use OpenGL and CoreGraphics for rendering,
CoreImage for image processing

Use Grand Dispatch Central (implements datalevel and task-level parallelism)

Profile your apps using Xcode profiler

Adv: Use Neon intrinsics & OpenGL ES shaders
Performance and power efficiency go together.

Using Accelerate Framework

#import <Accelerate/Accelerate.h>!
!
...!
cv::Mat src, dst;!
...!
!
[Link](cvRound([Link]*0.75), cvRound([Link]*0.75), [Link]());!
!
vImage_Buffer srcbuf = { [Link], [Link], [Link], [Link] };!
vImage_Buffer dstbuf = { [Link], [Link], [Link], [Link] };!
!
vImageScale_Planar8( &srcbuf, &dstbuf, NULL, kvImageNoFlags ); !

Formats: {ARGB, RGBA, Grayscale} x {byte, float}

Functionality: geometrical transforms, non-separable linear filters,
morphology, depth conversions, histogram equalization, LUT
Likely, Neon-based implementation (no GPU?)
For even better results use CoreImage and OpenGL ES
Accelerate also includes veclib, vDSP, BLAS.

Resources

[Link] (including excellent

introductory guides to Objective-C, Cocoa and
iOS programming)

[Link] (see iOS tutorial

to be greatly extended; we have dedicated GSoC
2012 project on OpenCV+iOS)

[Link]
- our GSoC iOS project repository where more
and more interesting stuff is being added