How to apply a 3D Model on detected face by Apple Vision "NO AR" - swift

Question

With iPhoneX True-Depth camera its possible to get the 3D Coordinates of any object and use that information to position and scale the object, but with older iPhones we don't have access to AR on front-face camera, what i've done so far was detecting the face using Apple Vison frame work and drawing some 2D paths around the face or landmarks.
i've made a SceneView and Applied that as front layer of My view with clear background, and beneath it is AVCaptureVideoPreviewLayer , after detecting the face my 3D Object appears on the screen but positioning and scaling it correctly according to the face boundingBox required unprojecting and other stuffs which i got stuck there, i've also tried converting the 2D BoundingBox to 3D using CATransform3D but i failed! i am wondering if what i want to achieve is even possible ? i remember SnapChat was doing this before ARKit was available on iPhone if i'm not wrong!
override func viewDidLoad() {
super.viewDidLoad()
self.view.addSubview(self.sceneView)
self.sceneView.frame = self.view.bounds
self.sceneView.backgroundColor = .clear
self.node = self.scene.rootNode.childNode(withName: "face",
recursively: true)!
}
fileprivate func updateFaceView(for result:
VNFaceObservation, twoDFace: Face2D) {
let box = convert(rect: result.boundingBox)
defer {
DispatchQueue.main.async {
self.faceView.setNeedsDisplay()
}
}
faceView.boundingBox = box
self.sceneView.scene?.rootNode.addChildNode(self.node)
let unprojectedBox = SCNVector3(box.origin.x, box.origin.y,
0.8)
let worldPoint = sceneView.unprojectPoint(unprojectedBox)
self.node.position = worldPoint
/* Here i have to to unprojecting
to convert the value from a 2D point to 3D point also
issue here. */
}

Answer 1

The only way to achieve this is to SceneKit with an ortographic camera and use SCNGeometrySource to match the landmarks from Vision to the vertices of the mesh.
First, you need the mesh with the same number of vertices of Vision (66 - 77 depending on which Vision Revision you're in). You can create one using a tool like Blender.
The mesh on Blender
Then, on code, on each time you process your landmarks, you do the steps:
1- Get the mesh vertices:
func getVertices() -> [SCNVector3]{
var result = [SCNVector3]()
let planeSources = shape!.geometry?.sources(for: SCNGeometrySource.Semantic.vertex)
if let planeSource = planeSources?.first {
let stride = planeSource.dataStride
let offset = planeSource.dataOffset
let componentsPerVector = planeSource.componentsPerVector
let bytesPerVector = componentsPerVector * planeSource.bytesPerComponent
let vectors = [SCNVector3](repeating: SCNVector3Zero, count: planeSource.vectorCount)
// [SCNVector3](count: planeSource.vectorCount, repeatedValue: SCNVector3Zero)
let vertices = vectors.enumerated().map({
(index: Int, element: SCNVector3) -> SCNVector3 in
var vectorData = [Float](repeating: 0, count: componentsPerVector)
let byteRange = NSMakeRange(index * stride + offset, bytesPerVector)
let data = planeSource.data
(data as NSData).getBytes(&vectorData, range: byteRange)
return SCNVector3( x: vectorData[0], y: vectorData[1], z: vectorData[2])
})
result = vertices
}
return result
}
2- Unproject each landmark captured by Vision and keep them in a SCNVector3 array:
let unprojectedLandmark = sceneView.unprojectPoint( SCNVector3(landmarks[i].x + (landmarks[i].x,landmarks[i].y,0))
3- Modify the geometry using the new vertices:
func reshapeGeometry( _ vertices: [SCNVector3] ){
let source = SCNGeometrySource(vertices: vertices)
var newSources = [SCNGeometrySource]()
newSources.append(source)
for source in shape!.geometry!.sources {
if (source.semantic != SCNGeometrySource.Semantic.vertex) {
newSources.append(source)
}
}
let geometry = SCNGeometry(sources: newSources, elements: shape!.geometry?.elements)
let material = shape!.geometry?.firstMaterial
shape!.geometry = geometry
shape!.geometry?.firstMaterial = material
}
I was able to do that and that was my method.
Hope this helps!

Answer 2

I would suggest looking at Google's AR Core products which support an Apple AR scene with the back or front facing camera...but adds some additional functionality beyond Apple, when it comes to non Face depth camera devices.
Apple's Core Vision is almost the same as Googles Core Vision framework which returns 2D points representing the eyes/mouth/nose etc...and a face tilt component.
However, if you want a way to simply apply either 2D textures to a responsive 3D face, or alternatively attach 3D models to points on the face then take a look at Google's Core Augmented Faces framework. It has great sample code on iOS and Android.