An empirical study on image bag generators for multi-instance learning

摘要

Multi-instance learning (MIL) has been widely used on diverse applications involving complicated data objects such as images, where people use a bag generator to represent an original data object as a bag of instances, and then employ MIL algorithms. Many powerful MIL algorithms have been developed during the past decades, but the bag generators have rarely been studied although they affect the performance seriously. Considering that MIL has been found particularly useful in image tasks, in this paper, we empirically study the utility of nine state-of-the-art image bag generators in the literature, i.e., Row, SB, SBN, k-meansSeg, Blobworld, WavSeg, JSEG-bag, LBP and SIFT. From the 6923 (9 bag generators, 7 learning algorithms, 4 patch sizes and 43 data sets) configurations of experiments we make two significant new observations: (1) Bag generators with a dense sampling strategy perform better than those with other strategies; (2) The standard MIL assumption of learning algorithms is not suitable for image classification tasks.