Domain Adaptation for Viewpoint Estimation with Image Generation

摘要

Viewpoint estimation is the pre-procedure for purposive perception and fine pose estimation, which has vital applications in robot manipulation and grasping. Unfortunately, the viewpoint estimation task is usually affected by a lack of training data with accurate annotations. In this paper, the problem of viewpoint estimation for untextured workpiece images is handled with the help of image translation and unsupervised domain adaptation. To augment the dataset, we transfer the 3D-rendering CAD images to the real-like images, which are visually similar to the real images, by utilizing a novel shape-constraint image generation model. Inspired by the ability of domain adaptation to bridge the domain gap, the distributions of these real-like images with real workpiece images are aligned by jointly optimizing the Maximum Mean Discrepancy (MMD) and H-divergence. Experiments show that our framework can achieve better authenticity for image translation, and improvements of the viewpoint estimation can be fulfilled over all the workpiece types, without the need of viewpoint labels of real workpiece images for training.