Learning to Predict 2D Object Instances by Applying Model-Based 6D Pose Estimation

Abstract

Object detection and pose estimation still are very challenging tasks for robots. One common problem for many processing pipelines is the big amount of object data, e.g. often it is not known beforehand how many objects and which object classes can occur in the surrounding environment of a robot. Especially available model-based object detection pipelines often focus on a few different object classes. However, new deep learning algorithms have been developed in the last years. They are able to handle a big amount of data and can easily distinguish between different object classes. The drawback is the high amount of training data needed. In general, both approaches have different advantages and disadvantages. Thus, this paper presents a new way to combine model-based 6D pose estimation with deep learning to reduce time for training and to improve the 6D pose estimation pipeline.