"... In PAGE 4: ... This uses as a semantic feature vector the vector of inner products between either a text query and each training label or test iamge and each training image. As shown in Tables 2 and 3 we compare the perfor- mance of the kernel-cca algorithm and generalised vector space model, where in Table2 we use a linear kernel as above for the image colour while in Table 3 we use a Gaus- sian kernel wtih AR BP max distance/20 . In both cases the kernel CCA method sharply outperforms GVSM.... ..."

"... In PAGE 4: ... This uses as a semantic feature vector the vector of inner products between either a text query and each training label or test iamge and each training image. As shown in Tables 2 and 3 we compare the perfor- mance of the kernel-cca algorithm and generalised vector space model, where in Table 2 we use a linear kernel as above for the image colour while in Table3 we use a Gaus- sian kernel wtih AR BP max distance/20 . In both cases the kernel CCA method sharply outperforms GVSM.... ..."

"... In PAGE 15: ... Our assumption is that the proposed method will outperform the conventional one in local photo semantic classification. Table3 shows the categorization results of the two different methods. The training and testing data was the same as the above experiment.... ..."

"... In PAGE 8: ... A yearly international workshop [33] [10] [17] [27] dedicated to the feature interac- tion (FI) problem and many other related publications provide strong evidence of the importance of the problem and the relevance to many researchers. Table5 summarizes works related to the FI problem and categorizes them according to the FM used.... ..."

"... In PAGE 8: ... A yearly international workshop [33] [10] [17] [27] dedicated to the feature interac- tion (FI) problem and many other related publications provide strong evidence of the importance of the problem and the relevance to many researchers. Table5 summarizes works related to the FI problem and categorizes them according to the FM used.... ..."

"... In PAGE 8: ... Ayearly international workshop #5B33#5D #5B10#5D #5B17#5D #5B27#5D dedicated to the feature interac- tion #28FI#29 problem and many other related publications provide strong evidence of the importance of the problem and the relevance to many researchers. Table5 summarizes works related to the FI problem and categorizes them according to the FM used.... ..."

"... In PAGE 5: ...nificant advantage of probability values for test cases, we consider this property to be less import- ant than simplicity and transparency for the exploratory analysis performed in this study, and have therefore chosen to base our analysis on classification tree results. The cross-validated errors of tree classification using a variety of features indicate that the combi- nation of SIFT with the Cb density is the most accu- rate of the feature sets considered ( Table4 ), with a balanced error of 20.5% and 29.... In PAGE 5: ...9.0 ^ 4.84 Classifiers were trained from the complete laboratory mutagenesis test set and used to predict the mutations in our human allele test set. Classification methods are the same as those described for Table4 , except that the cross-validation procedure was not used. Table 4.... ..."

"... In PAGE 7: ...e also experimented different variants of Tree Kernels, i.e. based on PTs. The results of the cross validation experiments are summarized in Table3 : Row 2 reports the results with polynomial kernel plus SST F (applied to linear features and a forest structure), Row 3 reports the cross validation results of polynomial kernel plus SST T (applied to linear features and a tree structure) and finally Row 4 illustrates the performance of the additive combination of polynomial with the PT kernel (PT T) (on linear features and a tree structure).... ..."