Seite - 3 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Figure1.Procedure forgeneratingannotated data, usinga robot andamovable fiducialmarker. portionsandmerged togaindatasetscontaining20% to100%(15k images)of all recordedsamples. The translational-15mm-errormetrics (percentage of tested data for which the translational error is smaller than 15 mm – accuracy necessary for grasp- ing)[7] inFigure2show, thatusingpre-trainedmod- els (blue, 6-10) leads to better performance than ini- tializing networks with random weights (red, 1-5). Bigger datasets do not necessarily improve the ac- curacy since biased datasets lead to wrong general- izations (e.g. network 5). A relatively small dataset recorded in the target domain achieves better results than a several times larger synthetic dataset (net- work 12: 15k real + 15k domain randomized im- ages), especially when extended using data augmen- tation (network 11: smallest real dataset augmented twice). Therotationalerrorsshowsimilar results,but are generally lower. Figure 2. Translational errors compared regarding train- ing time: Synthetic data (green), augmented data (cyan), pre-trained (blue) and non-pre-trained networks (red). Bubble-sizevisualizes dataset-size. A qualitative evaluation using a real mobile ma- nipulatorconfirms that theproposedpose-estimation system could be deployed in real-life logistics appli- cations to increase the levelof automation. References [1] U.Asif,M.Bennamoun,andF.A.Sohel. RGB-Dob- ject recognition and grasp detection using hierarchi- cal cascaded forests. IEEETransactionsonRobotics, 33(3):547–564, 2017. [2] G. Du, K. Wang, and S. Lian. Vision-based robotic grasping from object localization, pose estimation, graspdetection tomotionplanning: Areview.CoRR, 2019. [3] M.Garon,D.Laurendeau,andJ.F.Lalonde.Aframe- work for evaluating 6-DOF object trackers. In 15th EuropeanConference onComputer Vision – ECCV, pages 608–623,2018. [4] I. J. Goodfellow, Y. Bengio, and A. Courville. Deep Learning. MITPress,Cambridge,MA,USA, 2016. [5] P. Marion, P. R. Florence, L. Manuelli, and R. Tedrake. Label Fusion: A pipeline for generat- ing ground truth labels for real RGBD data of clut- tered scenes. In IEEE International Conference on RoboticsandAutomation– ICRA, pages 1–8,2017. [6] D.Pavlichenko,G.M.Garcı´a,S.Koo,andS.Behnke. Kittingbot: A mobile manipulation robot for collabo- rativekitting inautomotive logistics. In15th Interna- tionalConferenceonIntelligentAutonomousSystems – IAS, pages 849–864,2018. [7] J. Tremblay, T. To, B. Sundaralingam, Y. Xiang, D. Fox, and S. Birchfield. Deep object pose estima- tion for semantic robotic grasping of household ob- jects. In2ndAnnualConferenceonRobotLearning– CoRL, pages306–316,2018. [8] J.Walle´n. Thehistoryof the industrial robot. Techni- cal report fromAutomaticControl atLinko¨pingsuni- versitet, 2008. 3