RoCNet++: Triangle-based descriptor for accurate and robust point cloud registration

RoCNet++: Triangle-based descriptor for accurate and robust point cloud registration

2024 | Karim Slimani, Catherine Achard, Brahim Tamadazte
RoCNet++ is a novel point cloud registration method that introduces a triangle-based descriptor and a Farthest Sampling-guided Registration (FSR) module. The triangle-based descriptor encodes the local geometric properties of each point in the 3D point cloud by capturing the angles of triangles formed by the point and its nearest neighbors. This descriptor is designed to be robust to rigid transformations, as the angles of these triangles remain invariant under such transformations. The FSR module estimates the rigid transformation between two point clouds by computing multiple transformations on subsets of matched points, using a farthest point sampling strategy to ensure robustness and efficiency. The method was evaluated on three datasets: ModelNet40, KITTI, and 3DMatch. On ModelNet40, RoCNet++ demonstrated significant improvements in accuracy and robustness compared to other state-of-the-art methods, especially in noisy and partial overlap conditions. On KITTI, it achieved competitive performance in Registration Recall and Relative Rotation Error, while being the second-best in Relative Translation Error. On 3DMatch, RoCNet++ outperformed other methods in Inlier Ratio and Feature Matching Recall, and achieved comparable results in Registration Recall, Relative Translation Error, and Relative Rotation Error. The paper also includes an ablation study to validate the effectiveness of the triangle-based descriptor and the FSR module. The results show that the triangle-based descriptor significantly improves matching performance, and the FSR module enhances the accuracy of the estimated transformation. Future work will focus on improving noise robustness and extending the method to cross-modality registration and non-rigid correspondence estimation.RoCNet++ is a novel point cloud registration method that introduces a triangle-based descriptor and a Farthest Sampling-guided Registration (FSR) module. The triangle-based descriptor encodes the local geometric properties of each point in the 3D point cloud by capturing the angles of triangles formed by the point and its nearest neighbors. This descriptor is designed to be robust to rigid transformations, as the angles of these triangles remain invariant under such transformations. The FSR module estimates the rigid transformation between two point clouds by computing multiple transformations on subsets of matched points, using a farthest point sampling strategy to ensure robustness and efficiency. The method was evaluated on three datasets: ModelNet40, KITTI, and 3DMatch. On ModelNet40, RoCNet++ demonstrated significant improvements in accuracy and robustness compared to other state-of-the-art methods, especially in noisy and partial overlap conditions. On KITTI, it achieved competitive performance in Registration Recall and Relative Rotation Error, while being the second-best in Relative Translation Error. On 3DMatch, RoCNet++ outperformed other methods in Inlier Ratio and Feature Matching Recall, and achieved comparable results in Registration Recall, Relative Translation Error, and Relative Rotation Error. The paper also includes an ablation study to validate the effectiveness of the triangle-based descriptor and the FSR module. The results show that the triangle-based descriptor significantly improves matching performance, and the FSR module enhances the accuracy of the estimated transformation. Future work will focus on improving noise robustness and extending the method to cross-modality registration and non-rigid correspondence estimation.
Reach us at info@study.space
Understanding RoCNet%2B%2B%3A Triangle-based descriptor for accurate and robust point cloud registration