Point pair features are a popular representation for free form 3D objectdetection and pose estimation. In this paper, their performance in anindustrial random bin picking context is investigated. A new method to generaterepresentative synthetic datasets is proposed. This allows to investigate theinfluence of a high degree of clutter and the presence of self similarfeatures, which are typical to our application. We provide an overview ofsolutions proposed in literature and discuss their strengths and weaknesses. Asimple heuristic method to drastically reduce the computational complexity isintroduced, which results in improved robustness, speed and accuracy comparedto the naive approach.
展开▼
