Towards open-set semantic labeling in 3D point clouds : Analysis on the unknown class

Abstract

Abstract There has been a growing interest in the research of semantic labeling on scenes represented by 3D point clouds. A fundamental issue that has been largely ignored is the unavoidable presence of unknown objects and the lack of effective ways of dealing with them. Traditional methods usually label unknown objects as one of the pre-trained classes which is either a meaningful target class or a defined unknown class that collectively refers to all uninterested objects. Due to the fact that the class of unknown in essence is a collection of many unseen or uninterested classes, in which the in-class variation is significant and less manageable. It is challenging to solve the unknown problem in a pre-trained manner. In order to advance the research on semantic labeling with the presence of unknown objects, this study investigates the feasibility of adopting an open-set approach, i.e. train a model without unknown objects and reject them accurately in the test. In this paper, we propose a method that exploits the conflict of different labeling results in order to withstand the negative effect of unknown objects. The proposed framework relies on a Conditional Random Field (CRF) to capture inherent spatial relationships and appearance similarities between objects, and employs a Probability of Inclusion Support Vector Machine (PISVM) to estimate an unknown likelihood for each training class. The probabilistic outputs from both CRF and PISVM are then proposed to be combined under the Dempster Shafer theory for conflict measurement and unknown rejection. The novelty lies in that the method encodes both contextual constrains and unknown likelihood for performance enhancement. Comprehensive experimental results on publicly available data sets are presented to show the negative effects of unknown objects and the improvements on labeling accuracy achieved by the proposed method.