Detection of non-identical duplicate consumer photographs

Consumers often make more than one photograph of the same scene, creating non-identical "duplicates" and similar "non-duplicates". In Kodak's consumer photography database, 19% of the images fall into this category. Automatic detection of duplicates, therefore, is extremely useful in consumer applications. In this paper, first we develop a model of the problem and introduce a new classification of different types of duplicates. Then, we introduce a novel framework that automatically distinguishes between non-identical duplicate and very similar non-duplicate images. Our approach is based on a multiple strategy framework that combines our knowledge about the geometry of multiple views of the same scene, the extraction of low-level features, the detection of a limited number of semantic objects, and domain knowledge. The approach consists of three stages: (1) global alignment, (2) detection of change areas, and (3) local analysis of change areas. We present a novel and extensive image duplicate database (255 image pairs from 60 rolls from 54 real consumers, labeled by 10 other people). We analyze labeling subjectivity in detail and present experiments using our approach.