Learning to Represent Multiple Object Classes on a Continuous Viewsphere

Existing work on multi-class object detection usually does not cover the entire viewsphere of each class in a continuous way: object classes from different viewpoints are either discretized into a few sparse viewpoints [12],or treated as entirely separate object classes [20]. In the present work, we describe an approach to multi-class object detection which allows sharing parts between different viewpoints and several classes while also learning a dense representation for the entire viewsphere of each class. We describe three learning approaches with different part sharing strategies in order to reduce the computational complexity of the learnt representation. Our approach uses synthetic training data to achieve a dense viewsphere coverage which also allows to perform object class and 3D pose estimation on single images