Deep Learning based Iterative 6D Pose Estimation

Estimating the 6D object pose from a single RGB image is a challenging problem in computer vision. But it has various applications in robotics and virtual or augmented reality. Current solutions can be divided into two categories. The first one try to regress the pose parameters directly, but the estimation accuracy is usually not satisfied. The other one try to firstly extract local features and then estimate the pose through RANSAC based PnP algorithm. Since the network is designed for local feature extraction, it is not end-to-end trainable. Meanwhile, the RANSAC process is also time consuming with the presence of noise and outliers. A new architecture for object 6D pose estimation is proposed in this work. It is composed of correspondences extraction part, pose estimation part and the orthogonal iteration part. The first tw'o parts can be treated as an end-to-end trainable network, while the last two parts are used to find the optimal pose parameters without including the RANSAC process. Experimental results show
^:
that the architecture has persuasive advantage over the commonly applied 6D pose estimation solutions.