Facial anthropometry is the science of measuring the human face by identifying and marking landmark points on the subject (face). Measurements between the landmarks are then either taken manually by calipers or estimated through image processing techniques using the markers. Such processes are slow and intrusive, and require human intervention. The authors propose a marker-free approach for facial anthropometry using Microsoft Kinect, which captures a depth map as well as an RGB image of the scene.
The depth map from Kinect is used to segment the human figure from the background, and the segmented depth map is used to extract the human face from the RGB image. The FaceTracker application programming interface (API) is then applied to the RGB facial image to identify and localize 66 2D landmarks on the face. Thirteen of them are selected and transformed into the 3D coordinates using the intrinsic parameters of Kinect. The Euclidean distance between the set of selected pairs is calculated to extract 11 linear facial measurements in millimeters.
To test the accuracy of the approach, three methods--caliper (manual), Kinect 1 shot, and Kinect 100 shot--are compared for 36 healthy adults (29 males and seven females). In the third method, 11 facial measurements are taken 100 times and the average of the measured values is considered (54.5 percent of the measured distances were found to be accurate with respect to the caliper system).
The paper presents a novel technique for facial anthropometry that is non-intrusive (marker-free), operator-independent, and does not require the subjects to stay still for a long time. However, the accuracy and the speed of the system need to be significantly improved before it can be meaningfully deployed.
Anyone working in computer vision or machine intelligence in general, or biometrics or computational forensics in particular, will benefit directly from reading the paper.