BCFN: Bi-Directional Consistent Fusion Network for 6D Object Pose Estimation

In this work, we introduce BCFN, a Bi-directional Consistent Fusion Network designed for 6D pose estimation from a single RGB-D image. Our goal is to fully leverage the appearance information from the RGB image and the geometric information from the depth image. Previous studies have primarily utilized fused depth data to guide learning, overlooking the RGB branch. The core insight of our approach is to use both the RGB and depth branches to supervise the network. We have developed a dual-branch network that integrates appearance and geometric information. Subsequently, we propose pixel-level and vertex-level losses for each branch, respectively. Moreover, we introduce a consistent loss to enhance the network's convergence. We demonstrate the superiority of our method through experiments conducted on the LM-O dataset.