A Consideration on Scene Recognition Method by Bag of Object Semantics

Abstract

Recent year, scene recognition has been studied actively. Scene recognition has mainly two methods, the BoVW (bag of visual words) based method and the DNN (deep neural network) based method. However, in these methods, the component of the feature vector is represented as a local visual feature (visual words) so it is difficult to assign semantics to each component. Thus, we propose a novel scene recognition method, which is named ʻbag of object semantics.ʼOur proposed method has used the frequency of objects in a target scene as the feature vector. Each component of the vector corresponds to each object in a target scene, so our proposed method has high interpretability.