Joint Hierarchical Category Structure Learning and Large-Scale Image Classification

摘要

We investigate the scalable image classification problem with a large numberof categories. Hierarchical visual data structures are helpful for improvingthe efficiency and performance of large-scale multi-class classification. Wepropose a novel image classification method based on learning hierarchicalinter-class structures. Specifically, we first design a fast algorithm tocompute the similarity metric between categories, based on which a visual treeis constructed by hierarchical spectral clustering. Using the learned visualtree, a test sample label is efficiently predicted by searching for the bestpath over the entire tree. The proposed method is extensively evaluated on theILSVRC2010 and Caltech 256 benchmark datasets. Experimental results show thatour method obtains significantly better category hierarchies than otherstate-of-the-art visual tree-based methods and, therefore, much more accurateclassification.