Learning Stereopsis from Geometric Synthesis for 6D Object Pose Estimation
This work addresses the performance gap in monocular pose estimation for robotics and AR/VR applications, offering an incremental improvement over existing methods.
The paper tackles the problem of monocular 6D object pose estimation by proposing a method that uses a short baseline two-view setting and a 3D geometric volume to combine features from adjacent images, achieving state-of-the-art results with robustness in occlusion.
Current monocular-based 6D object pose estimation methods generally achieve less competitive results than RGBD-based methods, mostly due to the lack of 3D information. To make up this gap, this paper proposes a 3D geometric volume based pose estimation method with a short baseline two-view setting. By constructing a geometric volume in the 3D space, we combine the features from two adjacent images to the same 3D space. Then a network is trained to learn the distribution of the position of object keypoints in the volume, and a robust soft RANSAC solver is deployed to solve the pose in closed form. To balance accuracy and cost, we propose a coarse-to-fine framework to improve the performance in an iterative way. The experiments show that our method outperforms state-of-the-art monocular-based methods, and is robust in different objects and scenes, especially in serious occlusion situations.