The qualitative structure of objects and their spatial distribution, to a large extent, define an indoor human environment scene. This paper presents an approach for indoor scene similarity measurement based on the spatial characteristics and arrangement of the objects in the scene. For this purpose, two main sets of spatial features are computed, from single objects and object pairs. A Gaussian Mixture Model is applied both on the single object features and the object pair features, to learn object class models and relationships of the object pairs, respectively. Given an unknown scene, the object classes are predicted using the probabilistic framework on the learned object class models. From the predicted object classes, object pair features are extracted. A final scene similarity score is obtained using the learned probabilistic models of object pair relationships. Our method is tested on a real world 3D database of desk scenes, using a leave-one-out cross-validation framework. To evaluate the effect of varying conditions on the scene similarity score, we apply our method on mock scenes, generated by removing objects of different categories in the test scenes.
展开▼
