Focal segmentation for robust 6D object pose estimation

In the field of augmented reality, 6D pose estimation of rigid objects poses limitations and challenges. Most of the previous 6D pose estimation methods have trained deep neural networks to directly regress poses from input images or predict the 2D locations of 3D keypoints for pose estimation; thus, they are vulnerable to large occlusion. This study addresses the challenge of 6D pose estimation from a single RGB image under severe occlusion. A novel method is proposed that is based on PVNet but improves its performance. Similar to PVNet, our method regresses target object segments and pixel-wise direction vectors from an RGB image. Subsequently, the 2D locations of 3D keypoints are computed using the direction vectors of object pixels, and the 6D object pose is obtained using a PnP algorithm. However, accurate segmentation of object pixels is difficult, particularly under severe occlusion. To this end, a focal segmentation mechanism is proposed that ensures accurate complete segmentation of occluded objects. Extensive experiments on LINEMOD, LINEMOD-Occlusion datasets validate the effectiveness and superiority of our method. Our method improves the accuracy of PVNet by 1.09 and 5.14 on average in terms of the 2D reprojection error and ADD metric, respectively, without increasing the computational time.