CoPhIR Image Collection under the Microscope

The content-based photo image retrieval (CoPhIR) dataset is the largest available database of digital images with corresponding visual descriptors. It contains five MPEG-7 global descriptors extracted from more than 106 million images from Flickr photo-sharing system. In this paper, we analyze this dataset focusing on 1) efficiency of similarity-based indexing and searching and on 2) expressiveness of combination of the descriptors with respect to subjective perception of visual similarity. We treat the descriptors as metric spaces and then combine them into a multi-metric space. We analyze distance distributions of individual descriptors, measure intrinsic dimensionality of these datasets and statistically evaluate correlation between these descriptors. Further, we use two methods to assess subjective accuracy and satisfaction of similarity retrieval based on a combination of descriptors that is recommended for CoPhIR, and we compare these results on databases of 10 and 100 million CoPhIR images. Finally, we suggest, explore and evaluate two approaches to improve the accuracy: 1) applying logarithms in order to weaken influence of a single descriptor contribution if it deviates from the rest, and 2) the possibility of categorization of the dataset and identifying visual characteristics important for individual categories.