Published on March 21, 2008
Image and Multimedia Categorization: Image and Multimedia Categorization Gabriela Csurka, Eric Gaussier, François Pacull, Florent Perronnin, Jean-Michel Renders Xerox Research Centre Europe France Generic Visual Categorization (GVC): The pattern classification problem which consists in assigning one or multiple labels to an image based on its semantic content Generic: cope with various object and scene types Challenge: handle variations in view, lighting, occlusion, and typical object and scene variations Generic Visual Categorization (GVC) Generic Visual Categorizer (GVC): Generic Visual Categorizer (GVC) “Bag-of-key-patches” inspired by the “bag-of-words” : Patch detection: ROI (Harris, segmentation, etc.), multi-scale grid, random. Feature extraction: SIFT, RIFT, shape context, moments, color, etc. Visual vocabulary: K-means, GMM + EM, mean-shift, etc. Histogram computation: accumulating hard or soft occurrences Classification: pLSA, SVM, logistic regression, etc. Universal and adapted vocabularies: Universal and adapted vocabularies F. Perronnin, C. Dance, G. Csurka and M. Bressan, Adapted Vocabularies for Generic Visual Categorization, ECCV 2006. Bi-partite histograms: For each image, compute one histogram per category on the combined vocabularies Separate the relevant information from the irrelevant one Bi-partite histograms F. Perronnin, C. Dance, G. Csurka and M. Bressan, Adapted Vocabularies for Generic Visual Categorization, ECCV 2006. Experimental results: Experimental results Among the best performing systems at Pascal VOC Challenge 3rd system, 2nd institution 5-fold CV results on a 35 categories database, about 40000 images: Collected for RevealThis, related to travel and leisure, only mono-labeled Computed equal error rate (EER) and mean classification error rate (MCER) Expected confusions influenced the poor classification results F. Perronnin, C. Dance, G. Csurka and M. Bressan, Adapted Vocabularies for Generic Visual Categorization, ECCV 2006. Multiple-view categorization: Multiple-view categorization Problem statement: Given a set of data annotated in two different set of categories (different view). An independently pre-trained categorizer system available for both set. The aim is to exploit possible dependencies between the two category system to refine the categorization decisions. Visual (GVC) Textual (PLC) J.-M. Renders, E. Gaussier, C Goutte, F. Pacull and G. Csurka, Categorization in Multiple Category Systems, ICML 2006. Asymmetric re-weighting: Asymmetric re-weighting From assuming leads to: where The potential: can be computed through MLE on some training data J.-M. Renders, E. Gaussier, C Goutte, F. Pacull and G. Csurka, Categorization in Multiple Category Systems, ICML 2006. Symmetric re-weighting: Symmetric re-weighting where From assuming we get from which J.-M. Renders, E. Gaussier, C Goutte, F. Pacull and G. Csurka, Categorization in Multiple Category Systems, ICML 2006. Text categorization: Text categorization Texts are modeled as co-occurrences of documents (d) and words (w) with associated classes (c). The models are learned using PLC: where are estimated through MLE The posterior class probabilities of a new data are given by: J.-M. Renders, E. Gaussier, C Goutte, F. Pacull and G. Csurka, Categorization in Multiple Category Systems, ICML 2006. Web page classification: Web page classification Dmoz (http://www.dmoz.org) directory of classified web pages: Considered documents related to Travel and Leisure Considered flat categories: e.g. Theme_Parks/Disney, Travelogues/Middle_East, Preparation/Passports_and_Visas, Transportation/Boating, Recreation/WaterActivities/Scuba_Diving, … For each document (2828) we extracted: Textual information Images corresponding to the root and first level of the linked pages Logos, buttons, too small images were automatically discarded Splitting the data in: 1068 documents containing about 9300 images 1760 “pure” text documents (used to train the PLC) J.-M. Renders, E. Gaussier, C Goutte, F. Pacull and G. Csurka, Categorization in Multiple Category Systems, ICML 2006. Experimental results: Experimental results All images in Dmoz were manually annotated (the mentioned 35 categories) For text we kept flat Dmoz categories (89) Results are means of 5 fold cross validation J.-M. Renders, E. Gaussier, C Goutte, F. Pacull and G. Csurka, Categorization in Multiple Category Systems, ICML 2006. Good and bad correlation example: Good and bad correlation example Label: Transportation/Boating Text: Boating Magazine Wins Awards in Boating Writers International the Annual Writing Contest Internet Directory Boat of the Year, The Sea Doo, Sea Ray Sundancer, Boat Test Instant Confidence Do sweat the small stuff Watch streaming videos online Finding Mr. Smoothie, Cat vs. V A scientific approach Party Boat, Play our web game … Label: Flora_and_Fauna\Birding Text: Bird watching on the Greek Island of Lesvos Lesvos is known to the birding world as one of the best locations in Europe to see migrating birds. The island s vast and varied landscape offers numerous species of birds exactly what they want and thus it is a haven not only for birds but for birdwatchers too who arrive in droves in the spring. This is also the best time to see the beautiful display of wildflowers. For people interested in birds the best place to stay is Skala …… J.-M. Renders, E. Gaussier, C Goutte, F. Pacull and G. Csurka, Categorization in Multiple Category Systems, ICML 2006. (i,j) in symmetric re-weighting : (i,j) in symmetric re-weighting Transportation/Train_and_RailRoads Flora_and Fauna/Underwater Boat Underwater Train SkyActivity Recreation/SkyActivities/Skydiving Recreation/SkyActivities/Ballooning Paintings Recreation/WinterActivities/Ice_Skating Recreation/Fishing Transportation/Limousines_and_Shuttles Transportation/Offroad_Vehicles Fishing Vehicle Travel/Preparation J.-M. Renders, E. Gaussier, C Goutte, F. Pacull and G. Csurka, Categorization in Multiple Category Systems, ICML 2006. Categorization of video data: Categorization of video data Video segments Cross-media Categorization Text Images Video Segmentation Analysis Keyframe Categorization Image Processing SHOTS ASR + Audio Segmentation Analysis Speech Processing Speaker Identification and Clustering TRANSCRIPTIONS Text Segmentation Analysis Text Processing Text Categorization STORIES Collaborative work within RevealThis EU Project. Categorizer’s improvement example: Categorizer’s improvement example Keyframes Transcript (manual) Cross Media Categorization Results: Cross Media Categorization Results 2 videos about Greece was cut in 189 video segments and annotated: based on their multimedia content (40 predefined categories) by two independent annotator groups Evaluated the automatic annotations with these annotations as ground truth For text considered only those categories related to text categories (31) For images considered only those categories related to image categories (33) Conclusion: Conclusion Generic Visual Categorization (ECCV 2006) Novel approach based on a universal vocabulary and class-specific vocabularies obtained through adaptation Main advantages: low cost, no parameter tuning required no bounding boxes or manual segmentation required scales with the number of categories Multiple View Categorization (ICML 2006) Proposed two re-weighted approaches to handle multiple-view categories Experiments shows that these approaches manage to capture correlation between views, even in the case of noisy training data. Their main advantages are: the models are very simple and be estimated directly from some training data no need for huge amount of training data can be used on top of any existing pre-trained categorizer (without retraining them)
Jesus Feeds 5000 Spiritual-Inspirational presentation. ... Published on August 16, 2007. Author: VolteMort. Source: authorstream.com
wcor06. Education. Tweet. 18. 03. 2008 0 views FP7 MTAPU. Travel-Nature. Tweet. 16. 08. 2007 0 views apostles symbols. Entertainment. Tweet. 28. 11. 2007 0 ...
处理 = 120;设置规则（ 什么是分频点？） Pro= cessor Setting Fundamentals . 长期以来， 人们ë= 45;分频器有一些错误的= 认识 ...