Multi-core Parallelization of Point Set Dissimilarities for Accelerating the Comparison of Bags with Many Instances.

The representation of images under the Multiple Instance Learning (MIL) approach tends to generate large bags of instances, particularly when the so-called block decomposition is used as the instance generation strategy. This implies that the computations required to calculate dissimilarities between bags may become extremely expensive, even reaching highly prohibitive execution times. As a solution to this drawback, this paper presents a parallelization strategy of four point set dissimilarity measures for comparing images represented as large bags, allowing to perform a faster MIL-based image classification by accelerating the computation of the pairwise point set distances. The proposed strategy for calculating the point set dissimilarity between two bags distributes the instances of the first bag among the number of available processing units and, concurrently, calculates the selected dissimilarity measure for each partition of the first bag against all the instances of the second bag to be compared. The parallel strategy was experimentally tested for a problem of automatic visual inspection, showing that an acceleration of up to 23 times can be obtained when using 48 processing units on a machine with 24 physical cores.