Prior Geometry Guided Direct Regression Network for Monocular 6D Object Pose Estimation

Monocular 6D object pose estimation aims to estimate 6 degrees of freedom pose of known objects, gaining attention. Correspondence-based methods are the mainstream methods. They analyze the geometric information in 2D RGB images and establish 2D-3D correspondences to calculate 6D pose. However, pose estimation accuracy suffers from that 2D RGB images can not provide enough geometric information. To solve this problem, We propose a novel prior geometry guided direct regression network (PGDRN), which fully uses the prior geometric knowledge contained in given object models. Precisely, we extract the prior feature from the object model and concatenate the color feature extracted from 2D images to construct the prior-color feature, aggregating the prior and viewpoint-specific geometric information, making our method's accuracy and robustness. Experiments on two well-known LM-O and YCB-V datasets show that our method significantly outperforms state-of-the-art (SOTA) methods.