An Improved Pyramid Matching Kernel
The pyramid matching kernel (PMK) draws lots of researchers attentions for its linear computational complexity while still having state-of-the-art performance. However, as the feature dimension increases, the original PMK suffers from distortion factors that increase linearly with the feature dimension. This paper proposes a new method called dimension partition PMK (DP-PMK) which only increases little couples of the original PMKs computation time. But DP-PMK still catches up with other proposed strategies. The main idea of the method is to consistently divide the feature space into two subspaces while generating several levels. In each subspace of the level, the original pyramid matching is used. Then a weighted sum of every subspace at each level is made as the final measurement of similarity. Experiments on dataset Caltech-101 show its impressive performance: compared with other related algorithms which need hundreds of times of original computational time, DP-PMK needs only about 4-6 times of original computational time to obtain the same accuracy.
dimension partition bags of features SVM pyramid matching kernel function object recognition
Jun Zhang Guangzhou Zhao Hong Gu
College of Electric Engineering, Zhejiang University, Hangzhou, China, 310027
国际会议
无锡
英文
52-61
2010-09-17（万方平台首次上网日期，不代表论文的发表时间）