Cross-Language Evaluation Forum CLEF Workshop 2009 Poster Boaster Session CLEF 2009 Poster Boaster Session Features influencing retrieval in Persian Language Ljiljana Dolamic, Jacques Savoy (Computer Science Departement University of Neuchatel, Switzerland) Track(s): Ad-hoc@CLEF CLEF 2009 Poster Boaster Session Evaluation Features • Indexing strategy • Stopword list removal • Relevance assessments • Language features IR models • Okapi • DFR - PL2 • LM . • tf idf CLEF 2009 Poster Boaster Session Risk of not using Light Stemming: Official run UniNEpe3 Query Index Model Query Expansion Single MAP TD word PL2 none 0.4057 TD word PL2 Roc 10 docs / 50 terms 0.4466 TD TD TD word plural perstem Okapi Okapi PB2 Roc 5 docs / 50 terms Roc 5 docs / 70 terms idf 10 docs / 50 terms 0.4228 0.4311 0.4462 TD light PL2 Roc 5 docs / 20 terms 0.4718 CLEF 2009 Poster Boaster Session Comb MAP Z-Score 0.4663 Alicante at CLEF 2009 Robust-WSD Task Javi Fernández, Rubén Izquierdo and José M. Gómez (Department of Software and Computing Systems and El Taller Digital, from University of Alicante) Track(s): Robust WSD (Ad Hoc) CLEF 2009 CLEF 2009 Poster Boaster Session Poster Boaster Session Semantic classes (I) Goal: Explore the benefits of using semantic classes in IR systems WND (WordNet Domains) BLC20 (Basic Level Concepts) Results GMAP: +1.77% MAP: +0.64% CLEF 2009 Poster Boaster Session Query expansion re-ranking (II) Goal: Explore the benefits of combining known query expansion models with: WordNet synsets’ relations WordNet synsets’ distances Results GMAP: +9.97% CLEF 2009 Poster Boaster Session Thank you Feel free to make any questions See you at the poster session CLEF 2009 Poster Boaster Session CLEF topics and Wikipedia articles: Did it blend? Nuno Cardoso Linguateca Faculty of Sciences University of Lisbon Track: GikiCLEF CLEF 2009 Poster Boaster Session Motivation Using Wikipedia to answer GikiCLEF topics proved to be quite difficult. Is it the task or the systems? Wikipedia raw knowledge Did the systems mined Wikipedia conveniently, or they did not manage to scratch the surface? Are we aware of the difficulties in finding the right answers in Wikipedia? Knowledge mined by a GikiCLEF system Let’s analyse the Wikipedia collection and the GikiCLEF task. CLEF 2009 Poster Boaster Session Questions addressed in the poster In order to mine Wikipedia properly, we must have a better insight on what we are drilling here... Is English a pivot language? If I want to search for answers on an Italian-biased subject, do I stick with English, or give a try in Italian Wikipedia articles? Are languages self-sufficient enough to answer most of the GikiCLEF topics (monolingual), or having a multilingual system pays off? Where are the answers? In the text? In infoboxes? In tables? Are categories useful? CLEF 2009 Poster Boaster Session Are Passages Enough? The MIRACLE Team Participation in QA@CLEF2009 María Teresa Vicente-Díez, César de PabloSánchez, Paloma Martínez, Julián Moreno Schneider, Marta Garrote Salazar (MIRACLE Team) Universidad Carlos III de Madrid Track: monolingual Spanish task at ResPubliQA CLEF 2009 Poster Boaster Session Main Features 2 runs explore new domains, question types and multilingual experiments Adaptation to the new requirements Indexes redesign Passages fallback strategy Domain-specific rules Query generation adjustment Temporal management addition Acronyms mining functionality Ranking module tunning CLEF 2009 Poster Boaster Session Experiments. Pending Issues Results for submitted runs Question Analysis Linguistic Analysis Question Question Classification Information Retrieval Name Right Wrong Overall accuracy mira091eses 161 339 0.32 mira092eses 147 352 0.29 Answer Selection Query Generation Answer Filter Information Retrieval Passage Fallback Strategy Passage retrieval alternatives Re-ranking Filtering answers Collection Indexer Linguistic Analysis Answer Temporal-constraints integration Offline Operations Document Collection Answer selection helps? Document Index System evolution Configuration including answer filtering (BL+AF) CLEF 2009 Poster Boaster Session Domain independent Multilingual approach Evaluation The UPC experience in Qast 2009 Pere R. Comas and Jordi Turmo Technical University of Catalonia (UPC) QA – QAst Track CLEF 2009 Poster Boaster Session Contents System Architecture Description of Modules Partial Evaluation Insights on NE recognition on Speech QAst Evaluation Results CLEF 2009 Poster Boaster Session It has colorful charts CLEF 2009 Poster Boaster Session TIA-INAOE'S PARTICIPATION AT IMAGECLEF 2009 (LARGE-SCALE PHOTOANNOTATION) Hugo Jair Escalante LABORATORIO DE TECNOLOGÍAS DEL LENGUAJE hugojair@inaoep.mx Track: ImageCLEF Photo Annotation CLEF 2009 Poster Boaster Session INSTITUTO NACIONAL DE ASTROFÍSICA, ÓPTICA Y ELECTRÓNICA KNN-BASED IMAGE ANNOTATION A score is assigned to the labels from the KNN images to a test image: (liT ) = r wr (liT ) a wa (liT ) s ws (liT ) CLEF 2009 Poster Boaster Session LABELING REFINEMENT WITH COOCCURENCES The labels for the image are selected with an energy-based model that attempts to maximize the semantic cohesion of labels No_Visual_Season Outdoor Day No_Persons Partly_Blurred Neutral_Illumination CLEF 2009 Poster Boaster Session MIRACLE (FI) at ImageCLEFphoto 2009 R. Granados1, X. Benavent2, R. Agerri1, Ana García-Serrano3, J.M. Goñi1, J. Gomar2, E. de Ves2, J. Domingo2, G. Ayala2 1 Universidad Politécnica de Madrid, UPM 2 Universidad de Valencia 3 Universidad Nacional de Educación a Distancia, UNED Track: ImageCLEFphoto CLEF 2009 Poster Boaster Session Process Overview WORD (IDRA) TEXT-BASED RETRIEVAL PDF TXT Text Extractor XML Format + Selection PreProcess Collections/ captions BELGA CSV table INDEX Index VIEW IDRA Index Textual Results MySQL table MERGING (results list) Search Prepare Lists ENRICH NER tagger QUERIES file construction EQUImerge Generat e Lists Collections/ Topics CONTENT-BASED Collections / Imges RETRIEVAL (results list) BELGA (results list) CLEF 2009 Poster Boaster Session Merging algorithms (ImageCLEFphoto 2009) MirFI1 MirFI2 80 30 20 average best MirFI1 MirFI2 MirFI3 MirFI4 MirFI5 40 MirFI5 average MirFI4 MirFI3 50 % MirFI1 MirFI2 60 best 70 10 0 MAP R-Prec Metrics CLEF 2009 Poster Boaster Session CR@10 MirFI5 average 90 MirFI3 MirFI4 best CBIR module re-ranks obtained results from Textual one, using visual features extraction and comparation techniques. ENRICH enrichs a primary results list (Textual) using a support one (Visual) if a concrete result appears in both T and V lists, its relevance value is increased as follow: EQUImerge from a cluster-level results list, representatives results from each one of the different clusters are selected to build a new topic-level results list with diversity The University of Glasgow at ImageClefPhoto 2009 G. Zuccon, T. Leelanupab, A. Goyal, M. Halvey, P. Punitha, J. M. Jose Multimedia Information Retrieval, Uni. of Glasgow Track(s): ImageClef@CLEF CLEF 2009 CLEF 2009 Poster Boaster Session Poster Boaster Session Task: Promote Diversity in Image Ranking Textual diversity 1) MMR 2) Semantic Clustering 1.1) Tackle problem of non- optimal ranking due to initial bad document selection. 1.2) Incorporate clustering with MMR re-ranking 3) QPRP Visual diversity 4) FA and Bi-Clustering CLEF 2009 Poster Boaster Session Results Experimental results are very promising, where our approaches compared against state of the art. Our runs performance based on text diversification outperforms other participants’ based on text features only. CLEF 2009 Poster Boaster Session IAM@ImageCLEF 2009 Jonathon Hare, David Dupplaw and Paul Lewis Intelligence Agents Multimedia Group School of Electronics and Computer Science University of Southampton Track(s): ImageCLEF Photo Retrieval & Photo Annotation CLEF 2009 CLEF 2009 Poster Boaster Session Poster Boaster Session Photo Retrieval: Promote Diversity Challenge: Diversify image search results Research question: Can we use image features to diversify result sets? How? Start with a sound text retrieval methodology! Then add image features to re-order results Novel maximal-distance ranking function with powerful image features CLEF 2009 Poster Boaster Session Photo Retrieval: Promote Diversity Results: Text based retrieval worked rather well! Image-feature based diversification works But, at the cost of precision, with the current technique Future Work: Modify re-ranking algorithm to improve accuracy Build into a real retrieval system Thanks to: The EU for funding LivingKnowledge, which is investigating bias and diversity in search. CLEF 2009 Poster Boaster Session THE SAIAPR TC12 BENCHMARK http://imageclef.org/SIAPRdata Hugo Jair Escalante LABORATORIO DE TECNOLOGÍAS DEL LENGUAJE hugojair@inaoep.mx CLEF 2009 Poster Boaster Session INSTITUTO NACIONAL DE ASTROFÍSICA, ÓPTICA Y ELECTRÓNICA * In collaboration with Michael Grubinger EXTENDING THE IAPR TC12 BENCHMARK We manually segmented and annotated every image in the IAPR TC12 collection with the goal of providing ground-truth data for the evaluation of image annotation methods CLEF 2009 Poster Boaster Session THE RESOURCE The annotation vocabulary is organized into a conceptual hierarchy Landscape Sky Sky-blue Sky-light Sky-night Sky-red Cloud Ground Rock Sun Water Arctic Waterreflection Waves Glacier Ice Snow Lake Ocean River Swimmingpool Waterfall We provide: Beach Sand-beach Shore Island Vegetation Mountain Flowerbed Flower Fruit Apple Grapes Orange Strawberry Cactus Grass Mushroom Plant Leaf Trees Bush Palm Tree Branch Trunk segmentation masks for 20,000 images, labels 99,535 regions visual features for 99,535 regions spatial relationships among regions in the 20,000 images CLEF 2009 Poster Boaster Session Hill Volcano Evaluating Fusion Techniques at Different Domains at ImageCLEF Subtasks Sergio Navarro, Rafael Muñoz, Fernando Llopis (University of Alicante) Track(s): IMAGE@CLEF Photo Retrieval, WikipediaMM, Medical Retrieval CLEF 2009 CLEF 2009 Poster Boaster Session Poster Boaster Session Objective: To evaluate ... Multimodal Local Context Analysis (MMLCA) evaluation on new collections. Improved version of Multimodal Reranking TF-IDF (MMRR TF-IDF) Subquery Generation Module for improving diversity and precision of the results. Non-medical terms filtering: avoids the use of non medical terms returned by the local expansion technique used. CLEF 2009 Poster Boaster Session Concluding Results MMLCA confirmed its better performance for the two generic domain collections in comparison with other local expansion techniques. Subquery Generation Module demonstrated to be specially effective when it is combined with Probabilistic Relevance Feedback (PRF). (5th best textual run in Photo Retrieval task). A term coocurrence based technique like LCA does not work properly in medical domain collections. A deep understanding of the reasons could be a key for future improvements. CLEF 2009 Poster Boaster Session i-score, Image – Semantic and COntent based Retrieval system Giannis Boutsis, Theodore Kalamboukis Information Processing Laboratory Department of Informatics Athens University of Economics and Business Track(s): Image@CLEF CLEF 2009 Poster Boaster Session CLEF 2009 Poster Boaster Session ImageCLEF 2009 Removal of duplicate records (captions) 39,310 Discreate documents remained Stopword removal: Lucene + {caption, figure, image} Stemming : Porter’s algorithm ColorHistogram (0.3) ColorLayout (0.3) EdgeHistogram (1.0) 1 p visual _ SCORE (i) visual _ sim (i, ik ) p k 1 mixed _ SCORE (i) 0,8 * semantic _ SCORE (i) 0.2 * visual _ SCORE (i) iS CLEF 2009 Poster Boaster Session Concluding Remarks Visual and textual features should be used together for information retrieval whenever they are both available to allow optimal access to varied multimedia sources. CLEF 2009 Poster Boaster Session Topological Localization of Mobile Robots using Probabilistic Support Vector Classification Yan Gao, Yiqun Li Department of Computer Vision and Image Understanding, Institute for Infocomm Research, Singapore Track(s): RobotVision@CLEF CLEF 2009 Poster Boaster Session System Description Feature extraction Normalized Gaussian derivatives on the L component of the LAB color space (Lx,Ly,Lxx,Lyy,Lxy) Codebook k = 32 three-tier spatial pyramid The classifier -- Probabilistic estimate for multi-class support vector classification Handling the unknown class y=UK if pδ < T Exploiting continuity of the sequence Find the most frequent value in the past h frames CLEF 2009 Poster Boaster Session Experiments and Results By incorporating the prediction of the unknown class based on the probability estimates the performance improves from a score of 700 to 784, or 12%. By applying a smoothing procedure to exploit continuity of the sequence, the performance further improves from 784 to 884.5, or 12.8%. CLEF 2009 Poster Boaster Session University of Amsterdam’s Visual Concept Detection System Koen van de Sande Theo Gevers Arnold Smeulders Track: ImageCLEF Large-Scale Visual Concept Detection and Annotation CLEF 2009 CLEF 2009 Poster Boaster Session Poster Boaster Session Automatic concept annotation Given an image… … is some concept is present in that image? Examples: Flower Sunset Outdoor Single Person Landscape Sunny CLEF 2009 CLEF 2009 Poster Boaster Session Poster Boaster Session Ingredients Keypoints Annotated data Color Lowest equal error rate 40 out of 53 concepts Bag-of-visual-words Clouds CLEF 2009 Poster Boaster Session Sports are difficult Come see the poster! CLEF 2009 http://www.colordescriptors.com Poster BoasterVisit Session for color descriptor software Non-parametric Density Estimation Algorithms Ainhoa Llorente, Suzanne Little, Stefan Rüger Knowledge Media Institute The Open University, UK Track: ImageCLEF@CLEF Photo Annotation CLEF 2009 Poster Boaster Session Architecture CLEF 2009 Poster Boaster Session Results MMIS 33 2 1245434554581.txt: Weighted Global Features MMIS 33 2 1245586552541.txt: Keyword Correlation MMIS 33 2 1245611281967.txt: Normalized Google Distance MMIS 33 2 1245674693001.txt: Adapted LESK (WordNet) MMIS 33 2 1245601239738.txt: Wikipedia Link-based Measure CLEF 2009 Poster Boaster Session XRCE participation in Large Scale Visual Concept Detection and Annotation Task Gabriela Csurka and Yan Liu Xerox Research Center Europe 6 chemin de Maupertuis 38240 Meylan, France CLEF 2009 Poster Boaster Session Generic Visual Categorization (GVC) Patch detection Feature extraction +0.1 x = -1.5 … -0.5 Visual dictionary High level features Classification 100 90 80 70 60 50 40 30 20 10 10 20 30 40 50 60 70 80 90 100 Patch detection: regular grid extracted at multiple scales Feature extraction: SIFT and local color statistics Visual dictionary: Gaussian Mixture Models High level image representation: Fisher Vectors and Image GMMs Learning and Classification: Non linear discriminative classifiers with late (score) combination. Post-processing to ensure constraints of the ontology. CLEF 2009 Poster Boaster Session Results CLEF 2009 Poster Boaster Session SZTAKI @ ImageCLEF 2009 Bálint Daróczy joint work with András A. Benczú́r, Zsolt Fekete, Dávid Nemeskey, Istvá́n Petrá́s, Dá́vid Sikló́si, Zsuzsa Weiner Data Mining and Web Search Group Computer and Automation Research Institute Hungarian Academy of Sciences Text Based Image Retrieval • Unstructured text, lots of text data cleaning • Okapi BM25 • Thesaurus • KL dissimilarity to increase cluster recall • @WikiMM09 - text + Theasurus + image MAP 0.1699 - pure text MAP 0.1583 Image Based Retrieval and Classification • Segmentation (colour , shape and texture descriptors) • SIFT descriptor (PhotoAnn) • Logistic regression based on Fisher kernel (PhotoAnn) •@PhotoAnn09 EER 0.291718 AUC 0.773133 Thank you! UPMC/LIP6 at ImageCLEFannotation 2009: Large Scale Visual Concept Detection and Annotation Ali Fakeri-Tabrizi, Sabrina Tollari, Ludovic Denoyer, Patrick Gallinari (Université Pierre et Marie CURIE – Paris 6 UMR CNRS 7606 – LIP6, France) Track : ImageCLEF ANR-06-MDCA-002 AVEIR CLEF 2009 Poster Boaster Session VCDT – Visual Concept Detection Task Automatically extracting visual concepts Visual descriptors Segmentation HSV • Use the relations between concepts Detect the exclusive concepts Co-occurrence matrix : “day” or “night” Filter the classification scores using an optimism rate Support Vector Machine with a convenient loss function Baseline CLEF 2009 Poster Boaster Session VCDT – Visual Concept Detection Task Automatically extracting visual concepts Graph Classification Each node represents an image Each edge is weighted by the similarity between each pair of images Propagate the label of each node trough the graph We have also participated in the AVEIR consortium runs ANR-06-MDCA-002 AVEIR CLEF 2009 Poster Boaster Session Semantic Inter-media Image Retrieval in Photographic Collections Osama El Demerdash, Leila Kosseim & Sabine Bergler CLaC Lab, Concordia University, Montreal CLEF 2009 Poster Boaster Session Overview CLEF 2009 Poster Boaster Session Results Results on both IAPR-TC12 and Belga Precision increases with Recall till 100 ?!! CLEF 2009 Poster Boaster Session A particle-filter-based self-localization method using invariant features as visual information Jesús Martínez-Gómez Alejandro Jiménez-Picazo Ismael García-Varea SIMD group (Intelligent system and data mining) University of Castilla-La Mancha Track: RobotVision@ImageCLEF CLEF 2009 Poster Boaster Session SIMD Approach Obligatory track Multi-classifier using SIFT Each test image is matched with all the training images Additional image processing to detect natural landmarks Hough transform with lines and squares detection Optional track Particle-filter-based localization method using: SIFT for the update phase Training images are processed offline to extract their points Additional image processing Population initialization if the system becomes unstable The stability of the system is estimated Frames are not classified when the uncertainty about the robot’s pose is high The complete robot’s pose can be obtained CLEF 2009 Poster Boaster Session Results for the tracks Final Score Correctly Classified Misclassified Not Classified Position 684 10th 217 1st *** Obligatory Track *** 511.0 676 330 *** Optional Track *** 916.5 1072 311 CLEF 2009 Poster Boaster Session Patent Retrieval Experiments in the Context of the CLEF IP Track 2009 Daniela Becks, Christa WomserHacker, Thomas Mandl, Ralph Kölle (University of Hildesheim) Track(s): IP@CLEF CLEF 2009 Poster Boaster Session Questions to be investigated Does a statistical approach work well in the patent domain? How does the implemented stemmer influence the retrieval process? Do we need to remove stopwords? CLEF 2009 Poster Boaster Session Our approach System based on Lucene Monolingual runs (post runs) for German and English Simple retrieval approach Stopword removal Stemming Indexing Term queries CLEF 2009 Poster Boaster Session Search Log Analysis at the University of Sunderland Michael Oakes, Yan Xu (University of Sunderland) Track(s): LADS@CLEF CLEF 2009 Poster Boaster Session Search Engine based on Query Logs Instead of text documents being indexed according to their content, they are indexed according to the search terms previous users have used in finding them. Text Query Associated keywords Match Ranked Sessions Search Engine CLEF 2009 Poster Boaster Session Previously Downloaded URLs Search logs Automatic Language Identification Frequency of trigrams: sequences of 3 adjacent characters Most searches are in English Most searches are in only one language Usually the query language is the same as the interface language Latin is similar to Italian CLEF 2009 Poster Boaster Session Morphological acquisition by Formal Analogy Jean-François Lavallée and Philippe Langlais University of Montreal Track(s): Morpho Challenge@CLEF CLEF 2009 Poster Boaster Session Overview Task : Identify morphological structure of words in an unsupervised manner. Method: Use analogies to find morphologically related words. Analogy: unreadable is to read what undoable is to do Result: Good results for 4 out of 5 langages CLEF 2009 Poster Boaster Session Schema and results CLEF 2009 Poster Boaster Session PROMODES – A probabilistic generative model for word decomposition Sebastian Spiegler, Bruno Golénia, Peter Flach Computer Science Department University of Bristol, UK Track: Morpho Challenge 2009 CLEF 2009 Poster Boaster Session PROMODES – A probabilistic generative model for word decomposition Finding the best segmentation of a given word Competitive results in Morpho Challenge 2009 Objective Results Model Application CLEF 2009 Poster Boaster Session Based on hidden boundary positions and resulting letter transitions Languageindependent, for fusional and agglutinative languages Chemnitz Retrieval Group: Xtrieval Framework @ CLEF 2009 Jens Kürsten, Thomas Wilhelm, Maximilian Eibl (Chemnitz University of Technology) Track(s): TEL@CLEF, Grid@CLEF, VideoCLEF, ImageCLEFPhoto CLEF 2009 Poster Boaster Session Tracks & Technologies (1) TEL@CLEF Xtrieval framework with different IR cores Multilingual approach using multiple indexes Topic translation with Google’s Language API Grid@CLEF Implementation of CIRCO in Xtrieval Comparing and combining different IR models CLEF 2009 Poster Boaster Session Tracks & Technologies (2) VideoCLEF Classification as IR task with Xtrieval Fixing number of assigned Docs per Label ImageCLEFPhoto Simple approach without taking into account the diversity of the topic Xtrieval Framework using Lucene or Lemur CLEF 2009 Poster Boaster Session CACAO a Multilingual Interface to Library Catalogues Alessio Bosca, Luca Dini alessio.bosca@celi.it, dini@celi.it TEL@CLEF CLEF 2009 Poster Boaster Session What is CACAO? CACAO offers an innovative approach for accessing, understanding and navigating multilingual textual content in digital libraries and Online Public Access Catalogues (OPACs). Linguistic agnostic approach. Resources for different languages (dictionaries, thesauri,..) included via Web Services CLEF 2009 Poster Boaster Session CACAO Suite Different panels allows librarian to mange the different modules, view logs and perform off-line tasks/updates. CLIR FEATURES: Multi-words, Named Entities Recognition, Terms Disambiguation, Faceted Search.. CLEF 2009 Poster Boaster Session MultiMatch: Access to Cultural Heritage in a MultiModal way Giuseppe Amato, Pasquale Savino, Franca Debole ISTI- CNR (Italy) CLEF 2009 Poster Boaster Session MultiMatch: here it is MultiMatch is a Web search engine specialized in the cultural heritage domain: Queries can be expressed in multiple languages Multilingual and multimedia retrieval Access to multiple sources of information (e.g. museums, cultural institutions, CH educational sites, tourist information portals) CLEF 2009 Poster Boaster Session Web from: Museums Libraries Archives Newspapers News agencies Personal Pages Blogs Van Gogh Museum Museums Databases National Gallery Image Similarity Full Text Media Repository CLEF 2009 Poster Boaster Session Demo: Automatic Langauge Identification Using Trigram Frequencies Michael Oakes, Yan Xu (University of Sunderland) Track(s): LADS@CLEF CLEF 2009 Poster Boaster Session The 5 constituent trigrams in “_KATZE_” Trigram _KA KAT ATZ TZE English 1.23 E-5 4.72 E-7 4.72 E-7 3.49 E-5 1.70 E-5 1.63 E27 French 5.00 E-7 5.00 E-7 5.00 E-6 5.00 E-7 1.01 E-5 6.31 E32 German 5.64 E-4 1.41 E-4 1.41 E-4 1.41 E-4 1.12 E-4 1.77 E19 CLEF 2009 Poster Boaster Session ZE_ Overall Probability Automatic Language Identification Frequency of trigrams: sequences of 3 adjacent characters Most searches are in English Most searches are in only one language Usually the query language is the same as the interface language Latin is similar to Italian CLEF 2009 Poster Boaster Session The University of Glasgow at ImageClef 2009 Robot Vision Task Yue Feng, Martin Halvey and Joemon M. Jose Multimedia Information Retrieval, Uni. of Glasgow Track(s): ImageClef@CLEF CLEF 2009 Poster Boaster Session Task: determine topological location of mobile robots Proposed Approaches - combine 2 main methods, from coarse to fine procedures, plus illumination filter 1) POI matching in test and training frames to coarse determination 2) Rule based reasoning to refine results Two sets of data used for calculation 1) All frames in training and test set 2) First frame out of every five continuous frames Total three runs of submission CLEF 2009 Poster Boaster Session Results Our run ranked the first, relative to other participants’ Algorithm based on point matching, rule based reasoning, and illumination filters applied on every frames is effective to identify location of robot. CLEF 2009 Poster Boaster Session This is THE END Enjoy your lunch and See you in the Pool Rooms for The Poster Session CLEF 2009 Poster Boaster Session