Estimation of Subjects that are Distributed at a High Density from Non-pixel Information Applying NMF

We propose a method that estimates locations of subjects that has attracted crowd's attention with high accuracy from a large number of digital photographs. Recently, attempts that observe real world from a large number of data, assuming person as a sensor, have been very active. In the attempts, there are studies that try to estimate subjects attracting the crowd's attention in real time by quickly collecting a large number of photographs. The studies are focused on the tendency that a photograph is taken when a photographer comes upon an event that attracts its interest. Some of the proposed methods realize high availability by using only photographing, which includes information about location and azimuth of the camera and it is automatically embedded into photograph. Date size of photographing information is very small compared to that of pixel information, hence the proposed method reduce load on a communication infrastructure. However, there is a problem in accuracy when subjects are distributed at a high density. When there are many attractive subjects in a small region, the traditional works cannot find them because of their sequential search strategy. The proposed method applies non-negative matrix factorization (NMF) to subject's estimation, and the method is able to estimate subjects accurately even in a case that it is difficult with the conventional methods.