Nose tip detection using ellipsoid fitting for 2.5D partial face data

The existing problem of current nose tip detection research is to extract the nose from any viewpoint. In this paper, we propose a new method for localizing the nose tip by using an ellipsoid fitting technique. A 2.5D partial face dataset which contains partial face data, captured from any viewpoint between +/-45 degrees, is used for the experiments. This proposed method consists of three steps. First, the 2.5D facial data is fitted to an ellipsoid. Second, each point on the facial surface is projected onto a major axis to define the nose region on the facial surface. Based on its Euclidean distance from the estimated center of the ellipse to the projected point on the major axis, the nose region is defined. In the final step, the Mahalanobis distance is used to search for the nose tip according to the point which is farthest from the estimated ellipse center. To evaluate the performance of the proposed method, the face of each subject is captured in five different viewpoints between -45 degree and +45 degree. In total, 22 subjects take part and 486 samples are collected. From the experiments, it is found that the nose tip is correctly detected with 65.02% accuracy. The results show that the proposed method should be improved for higher performance and ellipsoid method is satisfactory for primary nose detection. Advantages of this method are that it is stable to face rotation and avoids constant parameter setting as a method used in Ref. [7].