The resolution of PARCOR images

Since the recognition rates are also influenced by the resolution of PARCOR images, we have performed recognition experiments with different resolutions by spatially reducing the input image sizes. In the experiments, we used HLAC features computed from the average image, the first order PARCOR image, and the second order PARCOR image. The recognition rates were estimated by leave one out method using 450 samples (50 samples $\times$ 9 categories).

The results are shown in Table 2. The achieved recognition rates were all more than 97%. This means that the proposed features are effective to classify these simple gestures. It is also noticed that very low resolution ( $16 \times 12$ pixels) is enough to discriminate these simple gestures.

 
 

Table 2: Recognition rates with different resolutions.

Resolution	$64 \times 48$	$32 \times 24$	$16 \times 12$
Rate (%)	97.6	98.4	98.9