Deep object 6-DoF pose estimation using instance segmentation

Pose estimation algorithms' goal is to find the position and the orientation of an object in pace, given only an image. This task may be complex, especially in an uncontrolled environment with several parameters that can vary, like the object texture, background or the lightning conditions. Most algorithms performing pose estimation use deep learning methods. However, it may be difficult to create dataset to train such kind of models. In this paper we developed a new algorithm robust to a high variability of conditions using instance segmentation of the image and trainable on a virtual dataset. This system performs semantic keypoints based pose estimation without considering background, lighting or texture changes on the object.