Figure 1. When a 3D reconstructor is trained using only a single
viewperobject,becauseofambiguityinthe3Dshapeofanobject,
itreconstructsashape whichonly fitsthe observedview andlooks
incorrect from unobserved viewpoints (upper). By introducing a
discriminator that learns prior knowledge of correct views, the re-
constructor is able to generate a shape that is viewed as reasonable
from any viewpoint (lower).