Skip to content

Latest commit

 

History

History
613 lines (459 loc) · 25.9 KB

README.md

File metadata and controls

613 lines (459 loc) · 25.9 KB

Shape Detection and Drawing in Augmented Reality(AR)

Detect the shape of drawing objects (classes - line, triangle, rectangle, pentagon, Hexagon, circle) and draw in Augmented Reality. Also mention the shape type along with drawing.

Like : "L" - Line, "T" - Triangle , "R" - Rectangle, "C" - Circle, "P" - Pentagon, "H" - Hexagon

Input Image :

img_2662

Edge Detected Image :

img_0225

Find Contour & Fill it for visualization :

img_0226

Create Scene graph :

img_4427d21af932-1

Example :

2 4

Basic Steps :

step 1 : Create a mlmodel for Edge Detection (Generic type)

step 2 : Take the image from ARFrame & idenfify edges using edge_detection.mlmodel

step 3 : Find out the Contours in the edge_detected image & calculate the approximation points using openCV.

step 4 : Figure the Shapes with it's image co-ordinates from Approximation points

step 5 : Map the image co-ordinates of shapes into AR-world co-ordinates

step 6 : Render Scene Graph

In the project, pod is not installed, So please do, "$pod install" before running the project.


Create an Edge Detection CoreML model


Original Caffe Model : http://vcl.ucsd.edu/hed/hed_pretrained_bsds.caffemodel

The Github project is : https://github.com/s9xie/hed

Download Edge_detection CoreML model(58MB) from : https://drive.google.com/drive/folders/0B0QC-w3ZqaT1ZEtpSG5HOE5VWEk which contains 6 different type of Outputs.

I am using the Side-out of original model (dsn3 output) only to reduce the space complexity.

(virtualenv2.7) C02QP68UG8WP:CoreML creation ashis.laha$ python mlmodel_converter.py

================= Starting Conversion from Caffe to CoreML ======================

Layer 0: Type: 'Input', Name: 'input'. Output(s): 'data'.

Ignoring batch size and retaining only the trailing 3 dimensions for conversion.

Layer 1: Type: 'Convolution', Name: 'conv1_1'. Input(s): 'data'. Output(s): 'conv1_1'.

Layer 2: Type: 'ReLU', Name: 'relu1_1'. Input(s): 'conv1_1'. Output(s): 'conv1_1'.

Layer 3: Type: 'Convolution', Name: 'conv1_2'. Input(s): 'conv1_1'. Output(s): 'conv1_2'.

Layer 4: Type: 'ReLU', Name: 'relu1_2'. Input(s): 'conv1_2'. Output(s): 'conv1_2'.

Layer 5: Type: 'Pooling', Name: 'pool1'. Input(s): 'conv1_2'. Output(s): 'pool1'.

Layer 6: Type: 'Convolution', Name: 'conv2_1'. Input(s): 'pool1'. Output(s): 'conv2_1'.

Layer 7: Type: 'ReLU', Name: 'relu2_1'. Input(s): 'conv2_1'. Output(s): 'conv2_1'.

Layer 8: Type: 'Convolution', Name: 'conv2_2'. Input(s): 'conv2_1'. Output(s): 'conv2_2'.

Layer 9: Type: 'ReLU', Name: 'relu2_2'. Input(s): 'conv2_2'. Output(s): 'conv2_2'.

Layer 10: Type: 'Pooling', Name: 'pool2'. Input(s): 'conv2_2'. Output(s): 'pool2'.

Layer 11: Type: 'Convolution', Name: 'conv3_1'. Input(s): 'pool2'. Output(s): 'conv3_1'.

Layer 12: Type: 'ReLU', Name: 'relu3_1'. Input(s): 'conv3_1'. Output(s): 'conv3_1'.

Layer 13: Type: 'Convolution', Name: 'conv3_2'. Input(s): 'conv3_1'. Output(s): 'conv3_2'.

Layer 14: Type: 'ReLU', Name: 'relu3_2'. Input(s): 'conv3_2'. Output(s): 'conv3_2'.

Layer 15: Type: 'Convolution', Name: 'conv3_3'. Input(s): 'conv3_2'. Output(s): 'conv3_3'.

Layer 16: Type: 'ReLU', Name: 'relu3_3'. Input(s): 'conv3_3'. Output(s): 'conv3_3'.

Layer 17: Type: 'Convolution', Name: 'score-dsn3'. Input(s): 'conv3_3'. Output(s): 'score-dsn3'.

Layer 18: Type: 'Deconvolution', Name: 'upsample_4'. Input(s): 'score-dsn3'. Output(s): 'score-dsn3-up'.

Layer 19: Type: 'Crop', Name: 'crop'. Input(s): 'score-dsn3-up', 'data'. Output(s): 'upscore-dsn3'.

================= Summary of the conversion: =================================== Detected input(s) and shape(s) (ignoring batch size):

'data' : 3, 500, 500

Network Input name(s): 'data'.

Network Output name(s): 'upscore-dsn3'.

input { name: "data" shortDescription: "Input image to be edge-detected. Must be exactly 500x500 pixels." type { imageType { width: 500 height: 500 colorSpace: BGR } } }

output { name: "upscore-dsn3" type { multiArrayType { dataType: DOUBLE } } }

metadata { shortDescription: "Holistically-Nested Edge Detection. https://github.com/s9xie/hed " author: "Original paper: Xie, Saining and Tu, Zhuowen. Caffe implementation: Yangqing Jia. CoreML port: Ashis Laha" license: "Unknown" }


Use the CoreML Model for detecting Edge of ARFrame captured Camera Image

img_2662

img_0225

Open CV framework added :

step 1 : create pod file with : pod 'OpenCV'

step 2 : Create a bridging header

Create an objective-c file from “Cocoa-touch class”
name it - OpenCVWrapper 
Xcode is smart and proposes to create a bridging header. Click on Create Bridging Header.

step 3 : Configure the bridging header ($project_name-Bridging-Header.h)

#import "OpenCVWrapper.h" in the bridging header 

step 4 : Change to Objective-c++

change from OpenCVWrapper.m to OpenCVWrapper.mm

step 5 : Importing opencv

#import <opencv2/opencv.hpp>
#import "OpenCVWrapper.h"
into OpenCVWrapper.mm file. 

NOTED : You will get ERROR : enum { NO, FEATHER, MULTI_BAND }; because of “NO” enum name. #import <opencv2/opencv.hpp> above all other imports will resolve the issue.

Step 6 : Write a test code

In OpenCVWrapper.h —> -(void) isOpenCVWorking;
In OpenCVWrapper.mm —>  @Implementation 

-(void) isOpenCVWorking {
	     NSLog(@"It's working");
 }

@end

AND call this from Swift class like :

let openCVWrapper = OpenCVWrapper()
	openCVWrapper.isOpenCVWorking()

It will generate Output : "It's working”

Finding Shape :

step1 : convert image into CV Matrix

+(cv::Mat)CVMatFromImage:(UIImage *)image {

CGColorSpaceRef colorSpace = CGImageGetColorSpace(image.CGImage);
size_t numberOfComponents = CGColorSpaceGetNumberOfComponents(colorSpace);
CGFloat cols = image.size.width;
CGFloat rows = image.size.height;

cv::Mat cvMat(rows, cols, CV_8UC4); // 8 bits per component, 4 channels
CGBitmapInfo bitmapInfo = kCGImageAlphaNoneSkipLast | kCGBitmapByteOrderDefault;

// check whether the UIImage is greyscale already
if (numberOfComponents == 1){
    cvMat = cv::Mat(rows, cols, CV_8UC1); // 8 bits per component, 1 channels
    bitmapInfo = kCGImageAlphaNone | kCGBitmapByteOrderDefault;
}

CGContextRef contextRef = CGBitmapContextCreate(cvMat.data,                 // Pointer to backing data
                                                cols,                       // Width of bitmap
                                                rows,                       // Height of bitmap
                                                8,                          // Bits per component
                                                cvMat.step[0],              // Bytes per row
                                                colorSpace,                 // Colorspace
                                                bitmapInfo);                // Bitmap info flags

CGContextDrawImage(contextRef, CGRectMake(0, 0, cols, rows), image.CGImage);
CGContextRelease(contextRef);

return cvMat;

}

step 2 : apply Morphology Transformations

step 3 : Find Contour from Image

step 4 : Calculate Approximate points from Contour

step 5 : Based on Approximation size, define the shape

step 6 : Retrieve the Positions (co-ordinates), Radius, Center of Circle & other shapes

-(cv::Mat) shapeDetection :(UIImage *)image { // image is the result of Edge detection, it's in gray scale.

/*
 // Convert to grayscale
 cv::Mat gray;
 cv::cvtColor(src, gray, CV_BGR2GRAY);
 // Convert to binary image using Canny
 cv::Mat bw;
 cv::Canny(gray, bw, 0, 50, 5);
 imageView.image = [UIImage fromCVMat:gray];
 */

cv::Mat cameraFeed =  [OpenCVWrapper CVMatFromImage:image];
std::vector< std::vector<cv::Point> > contours;
std::vector<cv::Vec4i> hierarchy;

// before applying contour finding, apply Morphology Transformations

// Closing the image (Method-1)
cv:: Mat bw2;
cv:: Mat erodedBW2;
cv:: Mat se = getStructuringElement(0, cv::Size(5,5));
cv::dilate(cameraFeed, bw2, se);
cv::erode(bw2, erodedBW2, se);

// Closing the image (Method-2)
cv::morphologyEx(cameraFeed, erodedBW2, cv::MORPH_CLOSE, se);

// Find contour
findContours( cameraFeed, contours, hierarchy, CV_RETR_EXTERNAL,  CV_CHAIN_APPROX_SIMPLE, cv::Point(0, 0));

bool objectFound = false;
if (hierarchy.size() > 0) {
    
    for (int index = 0; index >= 0; index = hierarchy[index][0]) {
        cv::Moments moment = moments((cv::Mat)contours[index]);
        double area = moment.m00;
        objectFound = (area > 100)? true : false;
    }
    //let user know you found an object
    if(objectFound ==true){
        for(int i=0; i < contours.size() ; i++) {
            cv::drawContours(cameraFeed,contours,i,cvScalar(80,255,255),CV_FILLED);
        }
    }
    
    // let's infer the shape from contours , calculate approx length of contours
    std::vector<cv::Point> approx;
    for(int i = 0; i < contours.size(); i++) {
        cv::approxPolyDP(cv::Mat(contours[i]), approx, cv::arcLength(cv::Mat(contours[i]), true)*0.02, true);
        
        // Skip small
        if (!(std::fabs(cv::contourArea(contours[i])) < 100)) { // && cv::isContourConvex(approx)
            
            printf("\n\n\n .......Area : %.0f\t", std::fabs(cv::contourArea(contours[i])));
            
            cv::Point2f center;
            float radius = 0.0;
            NSString * shape = @"";
            
            switch (approx.size()) {
                case 2: // line
                    printf("Line");
                    shape = @"line";
                case 3: // Triangle
                    printf("Triangle");
                    shape = @"triangle";
                    break;
                case 4: // Rectangle
                    printf("Rectangle");
                    shape = @"rectangle";
                    break;
                case 5: // Pentagon
                    printf("Pentagon");
                    shape = @"pentagon";
                    break;
                case 6: //Hexagon
                    printf("Hexagon");
                    shape = @"hexagon";
                    break;
                default: // circle
                    printf("circle \t");
                    shape = @"circle";
                    cv::minEnclosingCircle(cv::Mat(contours[i]), center, radius);
                    printf("Approx size : %ld , radius = %.1f",approx.size(),radius);
            }
            
            NSMutableArray * positions = [[NSMutableArray alloc] init];
            
            if ([shape isEqual:@"circle"]) {
                NSDictionary * dict = @{    @"radius":  [NSNumber numberWithFloat:radius],
                                            @"center.x":[NSNumber numberWithFloat:center.x],
                                            @"center.y":[NSNumber numberWithFloat:center.y]
                                            };
                [positions addObject:dict];
            }
            for (int j = 0; j < approx.size(); j++) {
                NSDictionary * dict = @{ @"x":[NSNumber numberWithInt:approx[j].x], @"y":[NSNumber numberWithInt:approx[j].y]};
                [positions addObject:dict];
            }
            [self.shapesResults addObject:@{shape:positions}]; // update the dictionary
        }
    }
}
return cameraFeed;

}

step 7 : Convert cv::Matrix into UIImage again

+(UIImage *)ImageFromCVMat:(cv::Mat)cvMat {
NSData *data = [NSData dataWithBytes:cvMat.data length:cvMat.elemSize()*cvMat.total()];

CGColorSpaceRef colorSpace;
CGBitmapInfo bitmapInfo;

if (cvMat.elemSize() == 1) {
    colorSpace = CGColorSpaceCreateDeviceGray();
    bitmapInfo = kCGImageAlphaNone | kCGBitmapByteOrderDefault;
} else {
    colorSpace = CGColorSpaceCreateDeviceRGB();
    bitmapInfo = kCGBitmapByteOrder32Little | (cvMat.elemSize() == 3? kCGImageAlphaNone : kCGImageAlphaNoneSkipFirst);
}

CGDataProviderRef provider = CGDataProviderCreateWithCFData((__bridge CFDataRef)data);

// Creating CGImage from cv::Mat
CGImageRef imageRef = CGImageCreate(
                                    cvMat.cols,                 //width
                                    cvMat.rows,                 //height
                                    8,                          //bits per component
                                    8 * cvMat.elemSize(),       //bits per pixel
                                    cvMat.step[0],              //bytesPerRow
                                    colorSpace,                 //colorspace
                                    bitmapInfo,                 //bitmap info
                                    provider,                   //CGDataProviderRef
                                    NULL,                       //decode
                                    false,                      //should interpolate
                                    kCGRenderingIntentDefault   //intent
                                    );

// Getting UIImage from CGImage
UIImage *finalImage = [UIImage imageWithCGImage:imageRef];
CGImageRelease(imageRef);
CGDataProviderRelease(provider);
CGColorSpaceRelease(colorSpace);

return finalImage;

}

Step 8 : Save the result for Visualization

img_0197 img_0196

cv::Mat cameraFeed = [self shapeDetection:image];
UIImage * result = [OpenCVWrapper ImageFromCVMat:cameraFeed];

// save it into photo-galary
UIImage * rotatedImage = [[UIImage alloc] initWithCGImage:[result CGImage] scale:1.0 orientation:UIImageOrientationRight];
UIImageWriteToSavedPhotosAlbum(rotatedImage, self, nil, nil);

Co-ordinate Mapping & SCNNode Create :

img_4427d21af932-1

step 1 : Create A Straight Line :

class func createline(from : SCNVector3 , to : SCNVector3) -> SCNNode { // Z is static
    // calculate Angle
    let dx = from.x - to.x
    let dy = (from.y - to.y)
    var theta = atan(Double(dy/dx))
    if theta == .nan {
        theta = 3.14159265358979 / 2 // 90 Degree
    }
    
    //Create Node
    let width = CGFloat(sqrt( dx*dx + dy*dy ))
    let height : CGFloat = 0.01
    let length : CGFloat = 0.08
    let chamferRadius : CGFloat = 0.01
    let route = SCNBox(width: width, height: height, length: length, chamferRadius: chamferRadius)
    route.firstMaterial?.diffuse.contents = UIColor.getRandomColor()
    let midPosition = SCNVector3Make((from.x+to.x)/2, (from.y+to.y)/2,0)
    let node = SCNNode(geometry: route)
    node.position = midPosition
    node.rotation = SCNVector4Make(0, 0, 1, Float(theta)) // along Z axis
    return node
}

step 2 : Create A Circle :

class func createCircle(center : SCNVector3, radius : CGFloat) -> SCNNode {
    var geometry : SCNGeometry!
    geometry = SCNCylinder(radius: radius, height: 0.01)
    geometry.firstMaterial?.diffuse.contents = UIColor.getRandomColor()
    geometry.firstMaterial?.specular.contents = UIColor.getRandomColor()
    let node = SCNNode(geometry: geometry)
    node.position = center
    node.rotation = SCNVector4Make(1, 0, 0, Float(Double.pi/2)) // along X axis
    return node
}

Step 3 : Create a Boundary :

class func boundaryNode() -> SCNNode {
    	let node = SCNNode()
    	let points : [(Float,Float)] = [(0.0,0.0),(0.5,0.0), (0.5,0.5), (0.0,0.5)]
    
    	for i in 0..<4 {
      	  	let x1 = points[i].0
        		let y1 = points[i].1
        		let x2 = points[(i+1)%points.count].0
        		let y2 = points[(i+1)%points.count].1
        
        		let from = SCNVector3Make(x1,y1,0)
        		let to = SCNVector3Make(x2,y2,0)
        		node.addChildNode(SceneNodeCreator.createline(from: from, to: to))
    	}
  	   return node
	}

Step 4 : Map from Image Co-ordinates into AR-Cordinates :

The Image Co-ordinates looks like :

.......Area : 9656	Triangle
 .......Area : 17871	Rectangle
 .......Area : 9368	circle 	Approx size : 8 , radius = 76.6
.......Area : 3100	Rectangle

Shape : triangle Values : (
    	{ x = 198; y = 255; },
    	{ x = 119; y = 373; },
    	{ x = 208; y = 485; })


Shape : rectangle Values : (
    { x = 303; y = 128; },
    { x = 231; y = 162; },
    { x = 247; y = 367; },
    { x = 330; y = 349; })


Shape : circle Values : (
    { "center.x" = 151; "center.y" = "106.5523"; radius = "76.61115"; },
    { x = 148; y = 30; },
    { x = 115;  y = 77; },
    { x = 112; y = 118; },
    { x = 127; y = 169; },
    { x = 156; y = 183; },
    { x = 183; y = 152; },
    { x = 191; y = 95;  },
    { x = 186; y = 60;  })


Shape : rectangle Values : (
    { x = 499; y = 0; },
    { x = 2; y = 0; },
    { x = 0; y = 499; },
    { x = 5; y = 8; })

The convertion function :

class func getSceneNode(shapreResults : [[String : Any]] ) -> SCNScene { // input is array of dictionary
    let scene = SCNScene()
    let convertionRatio : Float = 1000.0
    let imageWidth : Int = 499
    let xMin = 10
    let xMax = 490
    
    for eachShape in shapreResults {
        if let dictionary = eachShape.first {
            
            let values = dictionary.value as! [[String : Any]]
            switch dictionary.key {
            case "circle" :
                
                if let circleParams = values.first as? [String : Float] {
                    let x = circleParams["center.x"] ?? 0.0
                    let y = circleParams["center.y"] ?? 0.0
                    let radius = circleParams["radius"] ?? 0.0
                    let center = SCNVector3Make(Float(Float(imageWidth)-y)/convertionRatio+SceneNodeCreator.windowRoot.x, Float(Float(imageWidth)-x)/convertionRatio+SceneNodeCreator.windowRoot.y, SceneNodeCreator.z)
                    scene.rootNode.addChildNode(SceneNodeCreator.createCircle(center: center, radius: CGFloat(radius/convertionRatio)))
                    
                    // adding text
                    var textPosition = center
                    textPosition.y = textPosition.y + (radius/convertionRatio) + 0.01
                    scene.rootNode.addChildNode(SceneNodeCreator.create3DText("C", position: textPosition))
                    
                }
                
            case "line","triangle", "rectangle","pentagon","hexagon":
                for i in 0..<values.count { // connect all points usning straight lines (basic)
                    let x1 = values[i]["x"] as! Int
                    let y1 = values[i]["y"] as! Int
                    let x2 = values[(i+1)%values.count]["x"] as! Int
                    let y2 = values[(i+1)%values.count]["y"] as! Int
                    
                    // skip the boundary Rectangle here
                    if x1>xMin && x1<xMax {
                        let from = SCNVector3Make(Float(imageWidth-y1)/convertionRatio+SceneNodeCreator.windowRoot.x, Float(imageWidth-x1)/convertionRatio+SceneNodeCreator.windowRoot.y, SceneNodeCreator.z)
                        let to = SCNVector3Make(Float(imageWidth-y2)/convertionRatio+SceneNodeCreator.windowRoot.x, Float(imageWidth-x2)/convertionRatio+SceneNodeCreator.windowRoot.y, SceneNodeCreator.z)
                        scene.rootNode.addChildNode(SceneNodeCreator.createline(from: from, to: to))
                    }
                }
                
                // add shape description
                switch values.count {
                case 2: // line
                    let x1 = values[0]["x"] as! Int
                    let y1 = values[0]["y"] as! Int
                    let x2 = values[1]["x"] as! Int
                    let y2 = values[1]["y"] as! Int
                    
                    if x1>xMin && x1<xMax {
                        let center = SceneNodeCreator.center(diagonal_p1: (Float(x1),Float(y1)), diagonal_p2: (Float(x2),Float(y2)))
                        let centerVector = SCNVector3Make((Float(imageWidth)-center.1)/convertionRatio+SceneNodeCreator.windowRoot.x+0.01,
                                                          (Float(imageWidth)-center.0)/convertionRatio+SceneNodeCreator.windowRoot.y+0.01,
                                                          SceneNodeCreator.z)
                        scene.rootNode.addChildNode(SceneNodeCreator.create3DText("L", position: centerVector))
                    }
                    
                case 3 : // traingle
                    let x1 = values[0]["x"] as! Int
                    let y1 = values[0]["y"] as! Int
                    let x2 = values[1]["x"] as! Int
                    let y2 = values[1]["y"] as! Int
                    let x3 = values[2]["x"] as! Int
                    let y3 = values[2]["y"] as! Int
                    
                    if x1>xMin && x1<xMax {
                        let centroid = SceneNodeCreator.centroidOfTriangle(point1: (Float(x1),Float(y1)), point2: (Float(x2),Float(y2)), point3: (Float(x3),Float(y3)))
                        let centerVector = SCNVector3Make((Float(imageWidth)-centroid.1)/convertionRatio+SceneNodeCreator.windowRoot.x,
                                                          (Float(imageWidth)-centroid.0)/convertionRatio+SceneNodeCreator.windowRoot.y,
                                                          SceneNodeCreator.z)
                        scene.rootNode.addChildNode(SceneNodeCreator.create3DText("T", position: centerVector))
                    }
                    
                case 4: // Rectangle
                    let x1 = values[0]["x"] as! Int
                    let y1 = values[0]["y"] as! Int
                    let x2 = values[2]["x"] as! Int
                    let y2 = values[2]["y"] as! Int
                    if x1>xMin && x1<xMax {
                        let center = SceneNodeCreator.center(diagonal_p1: (Float(x1),Float(y1)), diagonal_p2: (Float(x2),Float(y2)))
                        let centerVector = SCNVector3Make((Float(imageWidth)-center.1)/convertionRatio+SceneNodeCreator.windowRoot.x,
                                                          (Float(imageWidth)-center.0)/convertionRatio+SceneNodeCreator.windowRoot.y,
                                                          SceneNodeCreator.z)
                        scene.rootNode.addChildNode(SceneNodeCreator.create3DText("R", position: centerVector))
                    }
                case 5,6: // pentagon, Hexagon
                    let x1 = values[0]["x"] as! Int
                    let y1 = values[0]["y"] as! Int
                    let x2 = values[3]["x"] as! Int
                    let y2 = values[3]["y"] as! Int
                    
                    if x1>xMin && x1<xMax {
                        let center = SceneNodeCreator.center(diagonal_p1: (Float(x1),Float(y1)), diagonal_p2: (Float(x2),Float(y2)))
                        let centerVector = SCNVector3Make((Float(imageWidth)-center.1)/convertionRatio+SceneNodeCreator.windowRoot.x,
                                                          (Float(imageWidth)-center.0)/convertionRatio+SceneNodeCreator.windowRoot.y,
                                                          SceneNodeCreator.z)
                        let text = (values.count == 5) ? "P" : "H"
                        scene.rootNode.addChildNode(SceneNodeCreator.create3DText(text, position: centerVector))
                    }
                default:
                    print("NO Shape")
                }
                
            default :
                print("This is default for Drawing node ")
            }
        }
    }
    // add boundary
    scene.rootNode.addChildNode(SceneNodeCreator.boundaryNode())
    return scene
}

2 4