Mixing deep learning with classical vision for object recognition

摘要

Nowadays, when one needs a system for image recognition, it is mostly a matter of finding pre-trained CNN and, sometimes, adding additional training based on transferred knowledge. Accurate 6-DOF object localization in the image is a more laborious task and requires more complex training data to be available. On the other hand, if we know the model of the object, it is straightforward to acquire its pose from the image (RGB or RGB-D). In this paper, we try to show the advantages of mixing deep learning object recognition/detection with classical 6-DOF pose estimation algorithms, with a focus on applications in service robotics.