Exploiting Temporal Information of Videos for Facial Landmark Detection: A Modified SDM

Facial landmarks are used for localization and representation of silent features of the face that are located on tips, corners and mid-points of facial segments. Detecting facial fiducial points in the video, play a fundamental role in facial images analysis during the past couple of decades due to its different facial image applications for example facial expression analysis, face recognition and gaze identification. Previous methods of face fiducial points detection gives good result of frontal face, but have problem generalizing to tilting and nodded faces. to solve this problem, we proposed a robust method for facial fiducial point detection for tilting and nodding faces in the video sequence using the second order regression based method. First, we detect the face from the first frame of the video sequence using Voila and Jones method. The proposed system extracts face from the videos and images and then detect the facial fiducial point using the Modified Supervised Descent Method (MSDM). Five public base databases YouTube celebrity, IMM, BU head tracking dataset, LFPW and VidTIMIT used to evaluate the performance. The comparison of proposed algorithm, i.e The proposed MSDM and existing algorithm SDM (Supervised Descent Method) has been made in terms of percentage accuracies. These accuracies are obtained by taking a distance measure of calculating facial landmarks with the ground truth. Results showed significant improvement and effectiveness against the standard datasets.