SIGGRAPH 2020 Papers

In this poster, we present a heterogeneous architecture for estimating 6D object pose from RGB images. First, we use a two-stream network to extract robust 3D-to-2D embedding feature correspondence. The segmentation stream processes the RGB information and spatial features individually. Then, we construct another fusion network to couple color and positional features, and predict the locations of keypoints in the regression stream. The pose can be obtained by an efficient RANSAC-based PnP algorithm. Moreover, we design an end-to-end iterative pose refinement procedure that further improves the reliable pose estimation. Our method outperforms state-of-the-art approaches in two public datasets.