Density-guided tree-structured kernel is proposed for the situation that the learning algorithm is not fit well for instances consisting of unordered feature sets with unequal cardinality.The feature set is automatically decomposed into a tree，and two feature sets are embedded into this tree to form two multi-resolution histograms，and then the histogram intersection，weighted by the density of feature points in common nodes from two multi-resolution histograms，is computed.The partial correspondences between feature sets can be determined automatically through this kernel，its computation is linear with the number of features，and it is positive-define.This kernel is embedded into kernel based discriminative classifier for image object classification，and compared with vocabulary-guided pyramid match kernel.The experiments show that the density-guided tree-structured kernel can obtain the better classification performance.