Adaptation of Synthetic Data for Coarse-to-Fine Viewpoint Refinement

Figure 1: Humans are perfect for annotating coarse viewpoints of objects in real images, but fail to estimate pose accurately at a fine level. 3D graphic models can be used to synthesize data at very accurate fine angles, but it is time-consuming to model all appearance variations present in real images. We therefore propose to leverage the abilities of humans of estimating coarse viewpoints and the pose accuracy of synthetic data.