Abstract:6D object pose estimation calculates the rotation and translation matrices from the object coordinate system to the camera coordinate system and plays an important role in tasks such as robotic grasping. The voting-based 6D pose estimation method PVNet votes on a set of hypotheses to determine one as the estimation for real keypoint, and uses Perspective-n-Point (PnP) algorithm to calculate 6D pose based on the estimated keypoints. For improving the accuracy of estimated keypoints, the accuracy of hypotheses should be improved firstly. Since each hypothesis is an intersection computed with extended lines of two predicted unit vectors, three factors should be considered for improving its accuracy. The deviation of angle between predicted vector and real vector should be as small as possible. The angular deviation for predicted vectors of pixels farther away from keypoints should be smaller than that of those nearer. Any two approximately parallel or coincident predicted vectors should be prohibited to compute intersection. In light of the three points, this paper predicts vector-field instead of unit vector-field to take into account the distance from pixel to real keypoint, and proposes a distance-aware vector-field prediction loss which requires that the farther pixels from keypoints, the smaller the angular deviation for predicted vectors, and suggests a strategy for preventing approximately parallel or coincident predicted vectors from computing hypothesis. Experiments on LINEMOD and OCC-LINEMOD datasets show that our method achieves 5.9% and 8.4% improvement for the average accuracy of pose estimation in terms of ADD(-S) respectively compared with PVNet.

Distance-Aware Vector-Field and Vector Screening Strategy for 6D Object Pose Estimation.

3D Point-to-Keypoint Voting Network for 6D Pose Estimation

PVNet: Pixel-Wise Voting Network for 6dof Object Pose Estimation.

KDFNet: Learning Keypoint Distance Field for 6D Object Pose Estimation

Attention Guided 6D Object Pose Estimation with Multi-constraints Voting Network

Efficient Encoding and Aligning Viewpoints for 6D Pose Estimation of Unseen Industrial Parts

PAV-Net: Point-wise Attention Keypoints Voting Network for Real-time 6D Object Pose Estimation

An Efficient Optimization Framework for 6D Pose Estimation in Terminal Vision Guidance of AUV Docking

3DPVNet: Patch-level 3D Hough Voting Network for 6D Pose Estimation

PVN3D: A Deep Point-wise 3D Keypoints Voting Network for 6DoF Pose Estimation

A 3D Keypoints Voting Network for 6DoF Pose Estimation in Indoor Scene

6D object pose estimation based on dense convolutional object center voting with improved accuracy and efficiency

A Robust CoS-PVNet Pose Estimation Network in Complex Scenarios

EP-Net: More Efficient Pose Estimation Network with the Classification-based Key-points Detection

WireframePose: Monocular 6D Pose Estimation of Metal Parts Based on Wireframe Extraction and Matching

VI-Net: Boosting Category-level 6D Object Pose Estimation via Learning Decoupled Rotations on the Spherical Representations

Adaptive Pose Estimation Algorithm Based on Point Pair Feature

Exploring Multiple Geometric Representations for 6DoF Object Pose Estimation

Prior-information-guided corresponding point regression network for 6D pose estimation

A Dynamic Keypoints Selection Network for 6DoF Pose Estimation

PANet: A Pixel-Level Attention Network for 6D Pose Estimation With Embedding Vector Features