An accurate box localization method based on rotated-RPN with weighted edge attention for bin picking

Abstract

Box localization that aims at localizing the position of the box plays a significant role in the application of bin picking. Accurate box localization is still a challenging problem. The boxes are stacked tightly with all kinds of angles, so that the border is very difficult to accurately locate. Moreover, the feature extraction used in existing methods are easily affected by the background, the shape, the size, the angle of box and the illumination. In this paper, we propose a novel Rotated Region Proposal Network (Rotated-RPN) wi attention to localize boxes in bin picking, named R-DFPN-WEA. It can generate inclined proposals with box orientation angle information and accurately localize the border of box. Moreover, we use point cloud information to extract the foreground information, getting rid of the background interference. Finally, the rotated masks with weighted edge attention are designed to enhance edges, which further improves the accuracy of box localization. We collect box images through depth cameras and preprocess the images, then create the box database, BoxLoc. Extensive experiments show that our proposed method, R-DFPN-WEA, achieves considerable improvement over the state-of-the-art approaches on our BoxLoc Datasets.