Intel RealSense SDK 2.0 for depth and tracking cameras

// Enable usage of librealsense frame data as a dlib image
rs_frame_image< dlib::rgb_pixel, RS2_FORMAT_RGB8 > image( color_frame );

// 1. Faces are detected in two steps:
// Facial boundary rectangles are detected:
  dlib::frontal_face_detector face_bbox_detector = dlib::get_frontal_face_detector();
  ...
  std::vector< dlib::rectangle > face_bboxes = face_bbox_detector( image );

// 2. Each one is then annotated to find its landmarks:
  dlib::shape_predictor face_landmark_annotator;
  dlib::deserialize( "shape_predictor_68_face_landmarks.dat" ) >> face_landmark_annotator;
  ...
  std::vector< dlib::full_object_detection > faces;
  for( auto const & bbox : face_bboxes )
    faces.push_back( face_landmark_annotator( image, bbox ));

if( nose_depth >= eye_depth ) return false;
if( eye_depth - nose_depth > 0.07f ) return false;
if( ear_depth <= eye_depth ) return false;
if( mouth_depth <= nose_depth ) return false;
if( mouth_depth > chin_depth ) return false;

// All the distances, collectively, should not span a range that makes no sense. 
// F.e. if the face accounts for more than 20cm of depth, or less than 2cm, then something's not valid
float x = std::max( { nose_depth, eye_depth, ear_depth, mouth_depth, chin_depth } );
float n = std::min( { nose_depth, eye_depth, ear_depth, mouth_depth, chin_depth } );
if( x - n > 0.20f ) return false;
if( x - n < 0.02f ) return false;

// Align depth and color frames
rs2::align align_to_color( RS2_STREAM_COLOR );
...
rs2::frameset data = pipe.wait_for_frames();
data = align_to_color.process( data );       // Replace with aligned frames