|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Competitions and international evaluation benchmarks
FGcomp'2013, in conjunction with Imagenet
Participants : Philippe-Henri Gosselin, Hervé Jégou.
Joint participation with Naila Murray and Florent Perronnin, Xerox Research Center Europe.
We have participated the the FGCOMP'2013 challenge and obtained the best results among all participants, see http://sites.google.com/site/fgcomp2013 Although the proposed system follows most of the standard Fisher classification pipeline, we have evaluated and used several key features and good practices that improve the accuracy when specifically considering fine-grained classification tasks [75] . In particular, we consider the late fusion of two systems both based on Fisher vectors, but that employ drastically different design choices that make them very complementary. Moreover, we show that a simple yet effective filtering strategy significantly boosts the performance for several class domains. The method is described in a technical report.
Hyperlink generation in broadcast videos
Participants : Guillaume Gravier, Pascale Sébillot, Anca-Roxana Simon.
Joint participation with Camille Guinaudeau, Heidelberg Institute of Technology (currently LIMSI-CNRS).
Following up on our 2012 participation, we participated in the Search and hyperlinking task implemented in the framework of the Mediaeval 2013 benchmark initiative. We limited ourselves to hyperlink generation, building on research results in natural language processing, information retrieval and topic segmentation, focusing our contribution on the selection of precise target segments for hyperlinks.
Maurdor campaign
Participant : Christian Raymond.
Joint participation with Yann Ricquebourg, Baptiste Poirriez, Aurélie Lemaitre and Bertrand Coüasnon, IRISA/Intuidoc.
We are participating to the ongoing MAURDOR campaign http://www.maurdor-campaign.org which aims at evaluating systems for automatic processing of written documents. The contribution of TexMex comes from the machine learning system based on boosting over bonsai trees we implemented. In the context of this campaign, we investigate the usefulness of this algorithm to combine efficiently features on a relatively big dataset. The very first result shows that this system get state-of-the-art performance while it is much faster than traditional SVM approaches.
Information extraction challenge at BioNLP-ST13
Participant : Vincent Claveau.
BioNLP Shared Task is a community-wide effort to address fine-grained, structural information extraction from biomedical literature. This year, several tasks were proposed and 22 teams participated. TexMex has proposed runs for three main tasks concerning entity extraction and categorization, and relation extraction. The methods proposed by our team are based on machine learning and information retrieval components. Although they do not exploit specialized or domain-specific knowledge, we obtained good results and ranked first, first and third according to the tasks.