A Higher-Fidelity Approach to Bridging the Simulation-Reality Gap for 3-D Object Classification

Computer vision tasks require collecting large volumes of data, which can be a time consuming effort. Automating the collection process with simulations speeds up the process, at the cost of the virtual data not closely matching the physical data. Building upon a previous attempt to bridge this gap, this thesis proposes three nuances to improve the correspondence between simulated and physical 3-D point clouds and depth images. First, the same CAD files used for simulated data acquisition are also used to 3-D print physical models used for physical data acquisition. Second, a new projection method is developed to make better use of all information provided by the depth camera. Finally, all projection parameters are unified to prevent the deep learning model from developing a dependence on intensity scaling. A convolutional neural network is trained on the simulated data and evaluated on the physical data to determine the model’s generalization ability.