Viewpoint selection for visual failure detection

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

Authors

• Akanksha Saran
• Branka Lakic
• Srinjoy Majumdar
• Jürgen Hess 0003
• Scott Niekum

Keywords

• Visualization
• Feature extraction
• Robot sensing systems
• Three-dimensional displays
• Support vector machines
• Failure Detection
• Viewpoint Selection
• Convolutional Neural Network
• Visual Features
• Task Execution
• Pre-trained Network
• Pre-trained Convolutional Neural Network
• Robotic Tasks
• Fine-grained Features
• Random Forest
• Image Features
• Classification Results
• Image Classification
• Highest Accuracy
• Small Datasets
• Average Accuracy
• Object Recognition
• ImageNet
• Selection Algorithm
• Random Forest Classifier
• Static Approach
• Deep Features
• Vision Sensors
• Results Of Task
• Convolutional Neural Network Features
• General View
• SVM Classifier
• Two-stage Selection
• Recovery Error
• 10-fold Cross-validation