Learning Category-level Implicit 3D Rotation Representations for 6D Pose Estimation from RGB Images

We exploit the embedding ability of a de-noising autoencoder for an implicit 3D rotation representation learning at the category level. Contrast to the exact 3D reconstruction model of each instance-level physical object, we leverage the inexact CAD/Reconstruction models of an object as the representative model for some category. Under our assumptions that objects within the same category share resemble geometry, we train a de-noising autoencoder on synthetic 3D views of category-level objects to extract the homogenous features at the bottleneck layer. The latent representation is agnostic not only to heterogeneous textures, colors, and illuminations, but also ambiguous pose caused by object symmetry. To extend the instance-level 3D translation estimation to the category level, we considered the 3D diagonal length ratio between the source and target object. We achieved a frame rate of 17Hz.