Recognizing Fine-Grained Object Instances for Robotics Applications

State-of-the-art object recognition systems, and computer vision methods in general, are getting better and better at a variety of tasks. Many of these techniques focus on general object categories, like person or bottle, and use many hundreds or thousands of training examples per category. Some domains, such as robotics, require recognition of fine-grained object instances, such as a 16oz. bottle of Coke, and only have access to limited training data. In this work we look at how to build data and adapt successful computer vision techniques for fine-grained object recognition tasks. Our Active Vision Dataset is the first deep-learning scale dataset for instance detection and motion simulation using real images. Our Target Driven Instance Detection method uses a siamese convolutional neural network to focus on specific fine-grained object instances, and has the capacity to generalize to instances outside its training set. Our SymGan method used the Generative Adversarial Network paradigm to train a 3D object orientation predictor that is robust to object symmetries.Doctor of Philosoph