Page - 129 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure 7. Comparison of predicted grasping candidates for both networks trained on automatically labeled data (top two rows) and manually labeled data (bottom two rows). We apply non-maximum suppression to reduce the number of visualizedboxesand toensure theclarityof thevisualization. Method Validgraspingcandidates in% Auto-Label 81.17 Man-Label 83.43 Table 1. Relative number of valid grasping candidates for both approaches. The network trained with automatically labeled data is named Auto-Label, whereas the network trained with manually labeled data is named Man-Label. Both networks show similar performance which empha- sizes theusefulness ofourautomatically labeleddata. Figure 8. Examples for non-graspable predictions. (Left) predictedboundingboxnotgraspablebecauseanotherob- ject ison top; (middle)boxtoobig; (right)boxtoosmall. 5.2.QualitativeResults Qualitative results of our grasping point predic- tions are shown in Figure 7 for the networks trained with the manually annotated data and the automati- callygenerated labels respectively. 6.Conclusion We have proposed an automatic annotation method for easily generating grasp proposals for robotic manipulations using only one RGBD cam- era. Ourannotationmethodrequiresminimalhuman interactionand ishighlycosteffective. With thepro- posed method, we generated ground truth data and successfully trainedadeepneuralnetwork topredict grasping candidates. To underline the usefulness of our approach, we trained our grasp prediction net- work with hand annotated and automatically anno- tated data separately, and our experiments showed similar performance for both attempts. This leads to the conclusion that our automatically generated la- bels arehighlyaccurate. We believe that the best strategy to train a deep net- work for grasping point predictions is to initially train with a large number of automatically annotated frames using our method, and afterwards fine-tune it with a small number of frames annotated by human experts. This strategy can lead to highly accurate re- sultswithminimalhuman interaction. 129