Few-Shot Viewpoint Estimation

two RexNet style networks for

• general keypoints
• 8 category specific semantic keypoints

then

• category specific viewpoint estimation block
• apply meta-learning
• estimate keypoints location, then solve equations for viewpoint

training

• MAML: min generalization loss each time
• keypoints’ 2d, 3d, depth location loss
• meta-siamese: nothing but an initializer?