Object relationship network model for image collections

Abstract

Image collections are indispensable parts of all image-based research. It is widely used for knowledge-training, comparing performance between systems, or verifying the effectiveness of the algorithms. Its role is not just source data of training, but it can affect performance and even can restrict focal planes of the image processing field. Various image datasets have appeared since PASCAL and Caltech-101, but there are few ways to verify the datasets themselves. This paper suggests the model that describes the collection of multiple images, based on the relationship between occurred objects in the images. Namely, object-relation network model for image collections. Recent image collections consist of dozens-GB data with a massive number of color pixels. It means it is practically impossible to compare the image collections directly, or even, to grasp properties of the individual collection. The model can work as single common criteria to evaluate the image collections.