A novel adaptive weighted fusion network based on pixel level feature importance for two-stage 6D pose estimation

In intelligent industry, accurate recognition and localization of objects in an image is the basis for robots to perform autonomous and intelligent operations. With the rapid development and application of deep learning data fusion technology in pose estimation, the existing 6D pose estimation methods have made many achievements. However, most of the existing methods are not accurate enough to cope with scenes with cluttered backgrounds, inconspicuous textures, and occluded objects. In addition, the existing methods ignore the effect of the accuracy of instance segmentation on the accuracy of pose estimation. To address above issues, this paper proposes a two-stage 6D pose estimation method based on adaptive pixel-importance weighted fusion network with lightweight instance segmentation, named TAPWFusion. In the instance segmentation stage, a lightweight instance segmentation network based on multiscale attention and boundary constraints, named CVi-BC-YOLO, is proposed to improve segmentation accuracy and efficiency. In the pose estimation stage, to eliminate the interference of lighting and occlusion, and enhance the accuracy of the pose estimation, we propose an adaptive pixel-importance weighted fusion network, named APWFusion, which adaptively evaluates the importance of RGB color and the geometrical information of the point cloud. Experiments on LineMOD, YCB-Video and T-LESS datasets prove the advanced and effective nature of our proposed method.