Deep Synthetic Viewpoint Prediction

  • Author / Creator
    Hess, Andy T
  • Determining the viewpoint (pose) of rigid objects in images is a classic vision problem with applications to robotic grasping, autonomous navigation, augmented reality, semantic SLAM and scene understanding in general. While most existing work is characterized by phrases such as "coarse pose estimation", alluding to their low accuracy and reliance on discrete classification approaches, modern applications increasingly demand full 3D continuous viewpoint at much higher accuracy and at real-time speeds. To this end, we here decouple localization and viewpoint prediction, often considered jointly, and focus on answering the question: How accurately can we predict, at real-time speeds, full 3D continuous viewpoint for rigid objects given that objects have been localized? Using vehicles as a case study, we train our model using only black and white, synthetic renders of fifteen cars and demonstrate its ability to accurately generalize the concept of "vehicle viewpoint" to color, real-world images of not just cars, but vehicles in general even in the presence of clutter and occlusion. We report detailed results on numerous datasets, some of which we have painstakingly annotated and one of which is new, in the hope of providing the community with new baselines for continuous 3D viewpoint prediction. We show that deep representations (from convolutional networks) can bridge the large divide between purely synthetic training data and real-world test data to achieve near state-of-the-art results in viewpoint prediction but at real-time speeds.

  • Subjects / Keywords
  • Graduation date
    Fall 2015
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.