Creating a new Dataset for Efficient Transfer Learning for 6D Pose Estimation
Australian National University, 2018
Citation: Henderson, J. (2018). "Creating a new Dataset for Efficient Transfer Learning for 6D Pose Estimation." Unpublished undergraduate Thesis, Australian National University.
Current state-of-the-art pose estimation techniques rely on neural network models to accurately detect, localise and estimate the pose of target objects in a scene. One of the primary weaknesses in these methods is the difficulty in detecting new target objects which the network has not been trained on. We propose a new type of neural network structure which allows for efficient transfer of learning from the training set onto unseen objects. We identify a lack of existing datasets to suit this type of structure. Based on this, we formalise a method for a semi-automated process of camera calibration, hand-eye calibration and capturing of images for the object pose library. We also consider methods of augmenting the pose library to create a corresponding training set for the target objects. We demonstrate how to isolate the target object from the background, and a method of artificially rotating the camera viewpoint. Overall, we provide a foundational process for creating a new type of dataset, enabling the use of a new structure of neural networks for pose estimation. We further identify a number of key areas for further development of this dataset.