Viewpoint selection for visual failure detection

摘要: The visual difference between outcomes in many robotics tasks is often subtle, such as the tip of a screw being near hole versus hole. Furthermore, these small differences are only observable from certain viewpoints or may even require information multiple to fully verify. We introduce and compare three approaches selecting for verifying successful execution tasks: (1) random forest-based method that discovers highly informative fine-grained features, (2) SVM models trained on features extracted pre-trained convolutional neural networks, (3) an active, hybrid approach uses above methods two-stage multi-viewpoint classification. These experimentally validated IKEA furniture assembly task quadrotor surveillance domain.