Fig 1.
(a) The template image and landmarks. (b) The patch representation of (a). A 9 × 9 patch is set around each landmark. (c) The target image with initial landmark. (d) The local matching results by NN. (e) The local matching results by 1-PP (KPP, K = 1). (f) The local matching results by 5-PP (KPP, K = 5). The size of the circle presents the weight value.
Fig 2.
Local matching search results of different methods including the nearest neighbor (NN) method, 1-PP (BBP), 2-PP, 4-PP and 6-PP method.
(a) and (b) show the average matching numbers when the initial location of landmarks are rotated 5 degrees and randomly perturbed by Gaussian noise N(15, 10), and rotated 10 degrees and randomly perturbed by Gaussian noise N(10, 10), respectively. (c) and (d) are the cumulative RMSE curves with corresponding initial locations of (a) and (b). The x-axis denotes the RMSE threshold and y-axis denotes the percentage of test images which are less than each threshold. The solid curves are the K-PP results without weight or the weights are equal to 1. The dashed curves are the K-PP results with weights.
Fig 3.
Comparison of the face alignment using LK, LK+NN and our methods on real images.
(a) Template with annotated landmarks. (b) Target image with initial shape parameter. (c) The results of Lucas-Kanade method. (d) The results of LK method with the nearest neighbor search. (e) The results of our method, which is Lucas-Kanade method with KPP search (K = 4).
Fig 4.
The cumulative RMSE curves with different initial parameters by four methods: Ours (LK+4-PP), LK+BBP, LK+NN and LK.
(a) The results with initial Gaussian noise N(0, 10). (b) The results with initial rotated N(0, 5) degrees and randomly perturbed by Gaussian noise N(15, 10). (c) The results with initial rotated N(0, 10) degrees and randomly perturbed by Gaussian noise N(10, 10).
Table 1.
The RMSE results for each facial component by different methods.