Predicting dynamic trajectories of pixels + surface normals under test action sequences

[yamins81]

with and without access to the action sequence input itself

[yamins81]

(1) how far into the future? [measured in frames
(2) using how much information about the past? [measured in frames]

[doug]

Instead. Let's start with surface normals as input, plus action and time and predict surface normal frame at timestep away. Thus learning a control nob for time.