Automatic Configuration of the Structure and Parameterization of Perception Pipelines

Abstract

The configuration of perception pipelines is a complex procedure that requires substantial amounts of engineering effort and knowledge. A pipeline consists of interconnected individual perception operators and their parameters, which leads to a large configuration space of pipeline structures and parameterizations. This configuration space has to be explored efficiently in order to find a solution that fulfills the specific requirements of the target application. In this paper, we present an approach to perform automatic configuration based on structure templates and sequential model-based optimization. The structure templates allow to reduce the search space and encode prior engineering knowledge. We introduce a structure template based on hypothesis generation, hypothesis refinement, and hypothesis testing to demonstrate the effectiveness of the approach. Experimental evaluation with state-of-the-art operators is performed on data from the T-LESS dataset as well as in a real-world robotic assembly task.