|
We intend to solve the problem of accurate facial alignment via a combined framework by using ASMs and AAMs. In general, ASMs are robust under various illumination conditions. However, the fitting accuracy needs improvement. AAMs have the potential to improve the accuracy although modeling appearance changes are challenging due to the variability of human faces. We have shown that AAMs can have a reasonable generalization power by increasing the size of training data explicitly. If the training size is too small compared to the dimension of the data, AAMs often tend to over-fit on the trained data, and thus might have poor generalization. In order to achieve the best combinations, ASMs and AAMs both need to produce reasonable performances respectively. The proposed CASAAMs, which compensates each other¡¯s weakness, can be applied in numerous applications, including pose estimation, expression analysis, shapefree face normalization, pose correction, and face recognition.
|
|
For reconstructing 3D faces, we propose a novel method of generating a 3D human face from a single image by utilizing an input face and an average depth-map. The point density of the input face and depth-map is increased simultaneously by the Loop subdivision method, which generates new vertices by a weighted sum of the existing vertices. Each point in an input image has an exact corresponding point in the depth-map whose intensity can be used for the estimation of depth in the input image. The computation time for our proposed 3D reconstruction method, together with feature detection, takes only 2-3 seconds, while that of a 3DMM takes 4-5 minutes with manual feature selection.conds.
|
