TAUbiomed_95_9_1246120389711 Use local patches histogram in multiple resolutions with an SVM classifier. Don't use wildcards and ignore code hierarchy. In the returned code replace trailing '0' with '*'. Idiap_3_9_1245417716666 LOW_lbp_sift_MULT_2MARG: The SVM algorithm makes the classification decision on the obtained margins for each image. Ordering them, if the first and the second highest margins differ minus than a treshold the corresponding two class labels are compared and where they differ we put a '*'. We applied this technique on the margins obtained from the LOW_lbp_sift_MULT run. 2005: treshold = 0.2 2006: treshold = 0.2 2007: treshold = 0.3 2008: treshold = 0.5 Idiap_3_9_1245417533955 LOW_lbp_sift_2MARG: The SVM algorithm makes the classification decision on the obtained margins for each image. Ordering them, if the first and the second highest margins differ minus than a treshold the corresponding two class labels are compared and where they differ we put a '*'. We applied this technique on the margins obtained from the LOW_lbp_sift run. 2005: treshold = 0.2 2006: treshold = 0.2 2007: treshold = 0.3 2008: treshold = 0.5 Idiap_3_9_1245417469975 LOW_lbp_sift: We combined the LBP features and the SIFT features at low level, that is simply concatenating them. The classification is done using SVM with the best parameters obtained from our validation experiments. Idiap_3_9_1245417671272 LOW_lbp_sift_MULT: We considered all the images belonging to classes with less than 10 elements and we multiplied them creating copies with modifications. The copies are added to the training set to enrich the corresponding classes. We applied this strategy on the low level combination of LBP and SIFT and we used SVM with the same best parameters obtained for the LOW_lbp_sift run. FEITIJS_96_9_1245937057229 We consider Edge Histogram Descriptors as global feature. As local descriptors we use SIFT descriptors and Local Binary Patterns. VPA_63_9_1245419336923 2005 and 2006 labels are obtained by the same single SVM classifier approach previously submitted. 2007 and 2008 labels, on the other hand, are recognized by hierarchical SVMs. Separate SVMs are trained on each subpart of the IRMA code for 2007 and 2008 labels. These individual results are then combined to form the resulting IRMA labels of our submission. SVM classifier hyper-parameters are C=100 and gamma=0,3. VPA_63_9_1245418900571 Single SVM classifier fed with LBP82 uniform features computed from sub-regions of images. Sub-regions are recovered such that each image is divided into 4x4 equal-sized, non-overlapping areas. SVM classifier hyper-parameters are C=100 and gamma=0,3. VPA_63_9_1245944101876 LBP features fed to hierarchical SVMs, which are organized with respect to data distribution. ClassSplit=10. C=100, g=0,3. Correct VPA_63_9_1246033855761 LBP features fed to hierarchical SVMs, which are organized with respect to data distribution. ClassSplit=10. C=100, g=0,3. More correct medGIFT_77_9_1245961041705 GIFT is used (Gray level = 8) together with aspect ratio information (10%). Descending weighting strategy and threshold are used. Results for 2007 and 2008 use chopping strategies. Database indexation : 2.5 hours Queries for 1000 similar images : 20 min Aspect ratio for all images: 10 min medGIFT_77_9_1245971471117 GIFT is used (Gray level = 16) together with aspect ratio information (10%). Descending weighting strategy and threshold are used. Results for 2007 and 2008 use chopping strategies. Database indexation : 2.5 hours Queries for 1000 similar images : 20 min Aspect ratio for all images: 10 min medGIFT_77_9_1246044416990 GIFT is used (Gray level = 16) together with aspect ratio information (10%). SIFT descriptor with bag of feature strategy is used and only common results appeared in both result lists are selected. Descending weighting strategy and threshold are used. Results for 2007 and 2008 use chopping strategies. VPA_63_9_1245936277557 LBP features fed to hierarchical SVMs, which are organized with respect to data distribution. ClassSplit=10. C=100, g=0,3. DEU_97_9_1246226037987 This method extracts global & local features including gray level co-occurrence matrix (GLCM) features, Mpeg-7 texture features. (Edge histogram descriptors). A k-Nearest Neighbor algorithm analyzes the extracted image feature vectors to determine the IRMA code associated to a given image. DEU_97_9_1246225040330 This method extracts global & local features including gray level co-occurrence matrix (GLCM) features, Mpeg-7 texture features. (Edge histogram descriptors).A k-Nearest Neighbor algorithm analyzes the extracted image feature vectors to determine the IRMA code associated to a given image. DEU_97_9_1245952497879 This method extracts global features including Gabor features, gray level co-occurrence matrix (GLCM) features, Mpeg-7 texture features. (Edge histogram, texture browsing, homogenous texture descriptors).A k-Nearest Neighbour algorithm analyzes the extracted image feature vectors to determine the IRMA code associated to a given image. DEU_97_9_1245952673253 This method extracts global Mpeg-7 texture features. (Edge histogram, texture browsing, homogenous texture descriptors).A k-Nearest Neighbour algorithm analyzes the extracted image feature vectors to determine the IRMA code associated to a given image.