6DoF Pose Estimation with Object Cutout based on a Deep Autoencoder

The six degree-of-freedom (6DoF) pose estimation is an important task in Augmented Reality, especially for initializing or recovering from the failure of 3D tracking for the textureless object, since it still encounters the insufficient accuracy problems because of cluttered backgrounds, occasionally quick movements, and other factors. We propose a simple but effective method to cutout the interested textureless object in a single RGB image with clear contour, which can be employed to directly estimate 6Dof poses with relatively high precision, and then help to remove most of the disturbing edges of clutter background for further refinement of pose estimation. To achieve this task, we propose a novel convolutional neural network, similar to an autoencoder, to reconstruct arbitrary scenes containing the object of interest, and extract the object area. We evaluated our method on objects from the LINEMOD dataset, and the results show that our approach is superior to the baseline as well as some advanced methods.