Here the initial representation is computed using an ICA algorithm. |
A toy dataset of two classes with eight points each | 1-d linear projection given by PCA/ICA/FDA |
The following figure shows the result using the ORL face dataset. As the plot shows, OCA achieves 100% recognization. Here the initial performance is given by an ICA algorithm.
The following figure shows the result using part of COIL 3D object dataset. As the plot shows, OCA achieves 100% recognization. Here the initial performance is given by an ICA algorithm.
The following figure shows the result using part of the CMU PIE dataset. As the plot shows, OCA achieves 100% recognization. Here the initial performance is given by an ICA algorithm.
The following figure compares OCA with other popular choices such as PCA, FDA, and ICA with respect to the number of subspaces on the ORL face dataset. The plots clearly show the effectiveness of OCA. Here the solid blue plot is the performance achieved by OCA with respect to the dimension of subspace; red dash-dotted ICA; green dashed FDA; black dotted plot PCA.