anooptomar1 / AppleFaceDetection

Face Detection with CoreML

Face Detection with Vision Framework

Previously, in iOS 10, to detect faces in a picture, you can use CIDetector (Apple) or Mobile Vision (Google)

In iOS11, Apple introduces CoreML. With the Vision Framework, it's much easier to detect faces in real time 😃

Try it out with real time face detection on your iPhone! 📱

You can find out the differences between CIDetector and Vison Framework down below.

Moving From Voila-Jones to Deep Learning

---

Details

Specify the VNRequest for face recognition, either VNDetectFaceRectanglesRequest or VNDetectFaceLandmarksRequest.

private var requests = [VNRequest]() // you can do mutiple requests at the same time

var faceDetectionRequest: VNRequest!
@IBAction func UpdateDetectionType(_ sender: UISegmentedControl) {
    // use segmentedControl to switch over VNRequest
    faceDetectionRequest = sender.selectedSegmentIndex == 0 ? VNDetectFaceRectanglesRequest(completionHandler: handleFaces) : VNDetectFaceLandmarksRequest(completionHandler: handleFaceLandmarks) 
}

Perform the requests every single frame. The image comes from camera via captureOutput(_:didOutput:from:), see AVCaptureVideoDataOutputSampleBufferDelegate

func captureOutput(_ output: AVCaptureOutput, didOutput sampleBuffer: CMSampleBuffer, from connection: AVCaptureConnection) {
    guard let pixelBuffer = CMSampleBufferGetImageBuffer(sampleBuffer),
        let exifOrientation = CGImagePropertyOrientation(rawValue: exifOrientationFromDeviceOrientation()) else { return }
    var requestOptions: [VNImageOption : Any] = [:]

    if let cameraIntrinsicData = CMGetAttachment(sampleBuffer, kCMSampleBufferAttachmentKey_CameraIntrinsicMatrix, nil) {
      requestOptions = [.cameraIntrinsics : cameraIntrinsicData]
    }
    
    // perform image request for face recognition
    let imageRequestHandler = VNImageRequestHandler(cvPixelBuffer: pixelBuffer, orientation: exifOrientation, options: requestOptions)

    do {
      try imageRequestHandler.perform(self.requests)
    }

    catch {
      print(error)
    }

}

Handle the return of your request, VNRequestCompletionHandler.

• handleFaces for VNDetectFaceRectanglesRequest
• handleFaceLandmarks for VNDetectFaceLandmarksRequest

then you will get the result from the request, which are VNFaceObservations. That's all you got from the Vision API

func handleFaces(request: VNRequest, error: Error?) {
    DispatchQueue.main.async {
        //perform all the UI updates on the main queue
        guard let results = request.results as? [VNFaceObservation] else { return }
        print("face count = \(results.count) ")
        self.previewView.removeMask()

        for face in results {
            self.previewView.drawFaceboundingBox(face: face)
        }
    }
}
    
func handleFaceLandmarks(request: VNRequest, error: Error?) {
    DispatchQueue.main.async {
        //perform all the UI updates on the main queue
        guard let results = request.results as? [VNFaceObservation] else { return }
        self.previewView.removeMask()
        for face in results {
            self.previewView.drawFaceWithLandmarks(face: face)
        }
    }
}

Lastly, DRAW corresponding location on the screen! <Hint: UIBezierPath to draw line for landmarks>

func drawFaceboundingBox(face : VNFaceObservation) {
    // The coordinates are normalized to the dimensions of the processed image, with the origin at the image's lower-left corner.

    let transform = CGAffineTransform(scaleX: 1, y: -1).translatedBy(x: 0, y: -frame.height)

    let scale = CGAffineTransform.identity.scaledBy(x: frame.width, y: frame.height)

    let facebounds = face.boundingBox.applying(scale).applying(transform)

    _ = createLayer(in: facebounds)

}

// Create a new layer drawing the bounding box
private func createLayer(in rect: CGRect) -> CAShapeLayer {

    let mask = CAShapeLayer()
    mask.frame = rect
    mask.cornerRadius = 10
    mask.opacity = 0.75
    mask.borderColor = UIColor.yellow.cgColor
    mask.borderWidth = 2.0

    maskLayer.append(mask)
    layer.insertSublayer(mask, at: 1)

    return mask
}