Image Recognition through Incremental Discriminative Common Vectors - Advanced Concepts for Intelligent Vision Systems - page 307

As with IDCV-EVD, the computation of Θ dominates the cost of the whole
algorithm, which will be referred to as IDCV-GSO. In this case, the cost of the
GSO-based DCV algorithm can be cut from O ( d ( M + N ) 2 )to O ( dN 2 )ifallnew
samples are linearly independent.

4 Experiments and Discussion

A number of experiments have been carried out to assess the relative benefits of
the IDCV algorithms with regard to the direct methods using data in a range of
situations. In this work, 3 publicly available image databases have been consid-
ered. Images were previously normalized in intensity, scaled and aligned using
the position of eyes and mouth. Figure 1 shows some sample images and their ba-
sic characteristics as dimensionality (image size), number of classes ( c ), number
of objects (per class), and type of variability. More details about these databases
can be found in the corresponding references also given as part of Figure 1.
In particular, an experimental setup in which more training data becomes
available to the algorithm has been designed. For each database, the available
data has been split into 3 disjoint sets. The first two are test (20%) and initial
training set (30%), respectively and the remaining 50% is made available as
new training data in portions of N images per class. Starting from a random
permutation of the images, test and train blocks in the partition have been
shifted throughout all the database so that all images have been used as test after
each evaluation round. Moreover, the incremental data subset has been randomly
permuted after each shift to remove any kind of dependence on the order in
which data is made available to the algorithm. The results presented correspond
then to an average across the whole database along with corresponding standard
deviations.

name [ref] image size c size/ c M N type

CMU-PIE [12] 120 × 160 68 56 [16,40] 4 expression, pose & light

JAFFE [13] 256 × 256 10 20 [6,14] 2 expression

Coil-20 [14] 128 × 128 20 72 [22,54] 4 pose

Fig. 1. Sample images from the 3 databases used in the experiments along with
corresponding details