Weakly-supervised Disentangling with Recurrent Transformations for 3D View Synthesis

observation

• human take time proportional to rotated angles in matching images of the same object
• recurrent: rotate 15 degrees once in a step

model

• encoder-decoder
• hidden unit: identity unit + pose unit
  • controlled by action unit
  • mimic the underlying manifold step by step
  • identity unit being shared while pose unit keeps changing
• also predict mask