DFW-PVNet: data field weighting based pixel-wise voting network for effective 6D pose estimation

Abstract

With the benefit of reduced memory and computational overhead, the sparse-based 6 degrees-of-freedom (6D) pose estimation method leverages the creation of sparse two-dimensional (2D) to three-dimensional (3D) correspondences to estimate the pose of objects in an RGB image. However, this method often leads to accuracy degradation. In this paper, we propose a data field weighting based pixel-wise voting network (DFW-PVNet), aiming at improving the accuracy of the 6D pose estimation while keeping excellent memory and computational overheads. The proposed DFW-PVNet first assigns potential weights to pixels at different positions by utilizing data field theory and then selects the pixels with higher potential weights to participate in the voting and locating of 2D keypoints. By building accurate sparse 2D-3D correspondences between the located 2D keypoints and the corresponding predefined 3D keypoints, the 6D pose of the object can be calculated through a perspective-n-point (PnP) solver. Experiments are conducted based on the LINEMOD and the Occlusion LINEMOD datasets, and the results show that the accuracy of the proposed method surpasses the state-of-the-art sparse-based methods and is comparable to dense-based methods but with significantly lower memory and computational overheads.