Estimating 6D Pose Using RANSAC Voting

Many recent works show that there is an extreme challenge for 6D pose estimation when objects are occluded from each other, especially since it is hard to infer invisible keypoints for some keypoint-based methods. In this paper, we propose a novel convolutional neural network for 6D pose estimation from a single RGB image. This network predicts the keypoints pixel by pixel using RANSAC voting. We evaluate this method on LINEMOD and Occlusion LINEMOD datasets, and generate our own datasets with several real objects for robotic grasping. The results show that the proposed network achieves a high accuracy for 6D pose estimation, and the grasping method based on our proposed network can grasp objects stably in real scenes.