CPF6D: 6DoF Pose Estimation Net based on Context Part Fusion for Robotic Grasping

Accurate 6D pose estimation of objects during robot grasping is a common task in robotics. It becomes even more challenging when the texture-less objects are situated in a complex environment. We present a part-level 6D pose estimation approach, CPF6D. It takes RGB-D images as input and predicts a 6D pose for each object in the image. This method is capable of integrating the relationships between the object itself and its constituent parts. In CPF6D, a feature extracting module is specifically designed in the encoder layer to enhance the contextual features of the object parts. We use a dual-stream fusion module for each RGB-D image to fuse the RGB and depth data. Also, an optimized strategy specifically for conveyor belt robotic grasping is presented. Experiments on synthetic datasets and in a real-world scenario demonstrate that our method can obtain accurate 6D pose of each object observed and be competitive with other public methods.