作者: Fengjun Lv , Shenghuo Zhu , Mert Dikmen , Yuanqing Lin , Thomas S. Huang
DOI:
关键词: Computer science 、 TRECVID 、 Software 、 Support vector machine 、 Search engine indexing 、 Artificial intelligence 、 Pattern recognition 、 NIST 、 Learning to rank 、 Contextual image classification 、 Coding (social sciences)
摘要: This notebook paper summarizes Team NEC-UIUC’s approaches for TRECVid 2010 Evaluation of Semantic Indexing. Our submissions mainly take advantage advanced image classification methods using linear coordinate coding (LCC) local features powered by the distributed computing software Hadoop. For every video shot, we evenly sample key frames and extract dense including DHOG LBP, which are encoded coding. Then, concept large-scale SVM classifiers trained based on spatial pyramid LCC features. Finally, employ multiple instance learning to rank shots according scores individual frames. systems achieve mean extended inferred average precision (mean xinfAP) 7.40% 30 concepts evaluated NIST 28.63% 1/5 development data as validation set total 130 concepts.