3D Context of Objects : A prior for Object Detection and Place Classification

Abstract: Contextual information is helpful for object detection and object-based placerepresentation. 3D data significantly helps to capture geometrical informationabout scenes. In this work, a feature descriptor for object context in full 3Dpointclouds of places is introduced together with a method to extract featuresand build the context model.The proposed model is evaluated in experiments on pointclouds from differenttypes of places which include different object categories. Results showthe promising ability of the model to predict the possible context of objects inpointclouds or complete 3D maps of an environment.Among various applications for this, the author suggests object contextmodels to be used in place categorization and semantic mapping and discussesa method for it. To the knowledge of the author, this work is unique regardingits use of full 3D pointcloud of scenes and also introducing this descriptor tobe used to represent places.