RotationNet Joint Object Categorization and Pose Estimation Using Multiviews from Unsupervised Viewpoints

method: treat veiwpoint as latent variables, learned unsupervised

• similar to frcnn, outputs prob of each category in each viewpoint, with a i_view stands for incorrect view
  • max. the prob. of each vp being the output vp and class label being target label -> EM!
    • determine vp: max. prob. of vp_i == v and vp_{j != i} = i_view
    • update params: max. log p., cross entropy optim.