Image clustering a personal clothing model

摘要

The disclosure is related to a system and method for learning robust clothing clustering based on a cluster ensemble technique applied to the clothing features of images to improve clustering of images. Different types of clothing features that are complementary to each other are computed to provide extensive description of the clothing in the images. Multiple partitions are computed based on the clothing features to generate a cluster ensemble set. A consensus function is applied to the multiple partitions to generate a final clothing consensus clustering that encompasses the information contained in the multiple partitions. A system and method are disclosed for clustering images based on the clothing of one or more persons in the images.

主权项

1. A method performed using a physical computer system comprising at least one processor for use in classifying images comprising:
generating, using the computer system, clothing feature data representative of at least one clothing feature using clothing region data representative of clothing regions detected in each image of a set of images; generating, using the computer system, a cluster ensemble set comprising a plurality of partitions, wherein each partition comprises a clustering of the images performed independently of the other partitions based on a clothing feature or a combination of clothing features, and the generating of the cluster ensemble set comprises
for each pair of clothing regions detected, computing a respective distance measure between the clothing feature data associated with the clothing regions to generate a distance matrix comprising the computed distance measures, andbased on the distance matrix computing partitions of the cluster ensemble set having different numbers of constituent clusters; and applying, using the computer system, a cluster consensus function to combine the plurality of partitions to generate a consensus clothing clustering of the set of images, wherein the consensus clothing clustering shares nonzero mutual information with each partition of the plurality of partitions.