@[TMM 2010] Real-Time Visual Concept Classification英文文献资料.pdf

1 Real-time Visual Concept Classification J.R.R. Uijlings, A.W.M. Smeulders and R.J.H. Scha Abstract—As datasets grow increasingly large in content based This paper presents a comprehensive evaluation of various image and video retrieval, computational efficiency of concept fast Bag-of-Words components in terms of both computa- classification is important. This paper reviews techniques to tional efficiency and retrieval performance, which is our main accelerate concept classification, where we show the trade-off between computational efficiency and accuracy. As a basis we contribution. The evaluation shows an increase of accuracy use the Bag-of-Words algorithm that in the 2008 benchmarks of for Random Forests [7] by combining them with Principal TRECVID and PASCAL lead to the best performance scores. We Component Analysis, and an increase in computational effi- divide the evaluation in three steps: (1) Descriptor Extraction, ciency by determining relevant image divisions for the Spatial where we evaluate SIFT, SURF, DAISY, and Semantic Textons. Pyramid [8]. Additionally, we present several improvements (2) Visual Word Assignment, where we compare a k-means visual vocabulary with a Random Forest and evaluate subsampling, upon the components under consideration: We provide a dimension reduction with PCA, and division strategies of the modified way to speed up the calculation of densely sampled 2 Spatial Pyramid. (3) Classification, where we evaluate the χ , SIFT [