Large Scale Visual Classification with Many Classes

• Published in: Machine Learning and Data Mining in Pattern Recognition pp. 629 - 643
• Main Authors: Doan, Thanh-Nghi, Do, Thanh-Nghi, Poulet, François
• Format: Book Chapter
• Language: English
• Published: Berlin, Heidelberg Springer Berlin Heidelberg
• Series: Lecture Notes in Computer Science
• Subjects: High Performance Computing; Large Scale Visual Classification; Parallel Support Vector Machines; Sampling Strategy
• ISBN: 3642397115; 9783642397110
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-642-39712-7_48

The usual frameworks for visual classification involve three steps: extracting features, building codebook and encoding features, and training classifiers. The current release of ImageNet dataset [1] with more than 14M images and 21K classes makes the problem of visual classification become more difficult to deal with. One of the most difficult tasks is to train a fast and accurate classifier. In this paper, we address this challenge by extending the state-of-the-art large scale classifier Power Mean SVM (PmSVM) proposed by Jianxin Wu [2] in two ways: (1) The first one is to build the balanced bagging classifiers with under-sampling strategy. Our algorithm avoids training on full data and the training process of PmSVM rapidly converges to the optimal solution, (2) The second one is to parallelize the training process of all classifiers with multi-core computers. We have developed the parallel versions of PmSVM based on high performance computing models. The evaluation on 1000 classes of ImageNet (ILSVRC 1000 [3]) shows that our approach is 90 times faster than the original implementation of PmSVM and 240 times faster than the state-of-the-art linear classifier (LIBLINEAR [4]).