“…Since it is hard to get all of the feature points when the absolute yaw and pitch angles are too large, we only selected 7466 images out of a total 15678 images so as to obtain the four feature points accurately by D-CNN [57] and AAM [58]. From Table IV, we can observe that Furthermore, in order to evaluate the performance of the proposed method comprehensively, we have compared the proposed method with some learning based methods (that need additional training processing and a large amount of training data), i.e., the improved Hough-voting with random forest [27], the DVFs and CNN-based method in [35], the fuzzy systems based method in [36], and the ellipsoidal model based geometry method in [42], and Kong's TIP2015 with AAM [49] in Table V for the Pointing'04 database, from which it can be observed that the MAE of the pitch and yaw angle of Proposed-MLFP is the smallest, while the MAE of the pitch and yaw angle of Proposed-ALFP with D-CNN is comparable to the start-of-the-art learning based methods.…”