Figure 8 Sample objects of the cellular phone database
Related Figures (19)
Fig. 1 Training process Fig. 2 Decision process For each database, we evaluated separately the classifica- tion error obtained using the First-Type Motion-Descriptors, the Second-Type Motion Descriptors, the Zernike Moments, and the combination of all three feature vectors. In this last case, the dimension of the feature space is d = 2 x 189+ 42 = 420. Fig. 4 Influence of number of training samples for COIL Table 1 Cross validated error rate on COIL-100 data-base Fig. 3. Sample objects of COIL-100 database Table 2 Error rate on COIL-100 noisy database Table 3 Error rate on ORL data-base Table 4 Error rate on A R face data-base Fig. 7 Face samples from the A R data-base Fig. 9 Influence of the number of training samples for the cellu- lar-phone data-base Fig. 10 Different lightings Fig. 11 Influence of contour-extraction on the number of training sam- ples with first motion-descriptors We also observed that the use of Second-type Motion- Notice the very important point that this correction term reserves the N log, N complexity of the FFT algorithm. For a function f constant over the cells of a regular grid it is easily computed that the exact correction term from the values of the FFT to values of the usual Fourier transform f at the points of the grid, is given by: Fig. 12 The vectors positions in the Fourier-space Fig. 13. Estimation of integrals in the discrete-Fourier-space prove the continuity of the quasi-representation of (Mo, ) that we construct in Sect. 3.4.5.
