In multi-view object detection, pose estimation should precede the decision stage of face or nonface. Pose estimation unit is implemented by designing a multi-class classifier of the predefined pose categories. Eigenspace approach is used throughout in this paper. For out-of-plane rotation estimation, Linear Discriminant Analysis (LDA) is employed in the proposed system, instead of the PCA reconstruction-based method. LDA provides a single linear subspace where data in each class may construct a cluster separable from the others. On the other hand, PCA uses different subspaces for the measurement of similarity to each eigenspace pose model, which sometimes gives vague criterion for classification. Moreover, the LDA-based algorithm is a lot faster since it only needs a single projection of input to the discriminant space, when PCA-based algorithm requires as many projections as the number of poses. One significant drawback in LDA is the lack of generalization capability, and this can be overcome by introducing Gabor response instead of normalized image. The benefit of applying Gabor response to the multi-view face detection problem is discussed in many aspects based on some experiments.
The validity of the proposed approach is verified through experiments, which include comparison among various combinations of proposed and previous approaches. And multi-view face detection results are given to show its effectiveness.