Anchor-Based 6D Object Pose Estimation

Estimating the 6D pose of known objects is an important task for many robotic and AR applications. The most recent trend in 6D pose estimation has been to take advantage of deep networks to either directly regress the pose from the image or to first predict the 2D locations of 3D keypoints and then recover the pose according to the 2D-3D coordinates relationship of keypoints using a PnP algorithm. However, most of these methods are vulnerable to occlusions. In this paper, we present a simple but efficient method to estimate 6D pose of objects using anchor-based corner detection, based on two-stage detection backbone (i.e. Faster R-CNN Ren et al. (2015)). Instead of directly predicting two-dimensional coordinates of the projected 3D bounding box (i.e. corners), we regress relative coordinates of top corners based on the two-dimensional anchor and diagonals of corresponding faces. Bottom corners are further robustly inferred using geometrical constraints of face diagonals. Experiments show that our method achieves significant improvement on LineMod S.Hinterstoisser et al. (2012) and Occlusion Brachmann et al. (2014a) dataset, outperforming most existing 6D pose estimation methods without using refinement.