An efficient network for category-level 6D object pose estimation

Abstract

Most category-level object pose estimation methods are multi-tasking, including instance segmentation, Normalized Object Coordinate Space (NOCS) map estimation and classification. However, previous approaches overlooked the connection between multiple tasks. In this work, we propose an efficient network to make better use of the complementarity between different tasks. Specifically, we propose an external sharing unit (ESU) to promote instance segmentation and NOCS map estimation. In addition, we propose an internal sharing unit (ISU) to improve the NOCS map estimation. The NOCS map head has three branches. And the estimated coordinates of each branch have strong correlation. Extensive experiments on the CAMERA and REAL dataset demonstrate the effectiveness of joint optimization in multi-tasking category-level object estimation. Experimental results also show that the proposed method can improve not only accuracy but also efficiency on several benchmarks.