Object detection constrained by ontological priors -

Abstract

The problem of object detection in Computer Vision is a difficult and interesting problem which is far from being solved due in no small part to the challenges of perception. Nevertheless, by introducing top-down priors such as semantics, the problem of segmenting and detecting objects becomes traceable. This paper proposes such an approach by relying on the ontological relationships that make up parts of objects in order to enhance their detection. The proposed method processes the point cloud of a scene and clusters it into pools of potential objects. Hypotheses on the object identity is generated using geometric and customized ontological definitions to generate probabilistic models that would constitute the building blocks for the decision making process. An object labeling scheme derived by minimizing an energy function is presented. Finally, objects are replaced by matching them to generic CAD models. To evaluate the proposed method, we run our experiments on three well-known datasets and compare with results in the literature. Results show superiority to the prior art in terms of both recall and precision.