VIREO at TRECVID 2010: Semantic Indexing, Known-Item Search, and Content-Based Copy Detection
Chong‐Wah Ngo, Shiai Zhu, Hung‐Khoon Tan, Wan‐Lei Zhao, Xiao-Yong Wei · 2010
This paper presents our approaches and the comparative analysis of our results for the three TRECVID 2010 tasks that we participated in: semantic indexing, known-item search and content-based copy detection. Semantic Indexing (SIN): Our main focus for the SIN task is on the study of the following two issues: 1) the effectiveness of concept detectors for indexing web video dataset, and 2) how to leverage the ontology relationships to reinforce concept detection. Our baseline detectors are similar to those of our TRECVID 2009 system, where both local and global features are employed to train the SVM model for each concept. Based upon the baseline detectors, we propose two approaches to refine the detection scores. The first integrates the ontology information into the random walk framework. The second seeks the agreement among the set of ranked lists generated by semantically related concepts. Our four submitted runs are summarized below:- F A VIREO.randomwalk 1: perform a random walk over the baseline result using local feature alone. Flickr distance and ontology relationship are used to build the context.- F A VIREO.agreement 2: re-rankthevideosbyseekingtheagreement amongsemantically related concepts on the relevant videos.- F A VIREO.baseline vk 3: local feature alone- multiple detectors.- F A VIREO.baseline vk cm 4: average fusion of local feature and global feature.