RPM-Net: Robust Point Matching using Learned Features (2003.13479v1)

Abstract: Iterative Closest Point (ICP) solves the rigid point cloud registration problem iteratively in two steps: (1) make hard assignments of spatially closest point correspondences, and then (2) find the least-squares rigid transformation. The hard assignments of closest point correspondences based on spatial distances are sensitive to the initial rigid transformation and noisy/outlier points, which often cause ICP to converge to wrong local minima. In this paper, we propose the RPM-Net -- a less sensitive to initialization and more robust deep learning-based approach for rigid point cloud registration. To this end, our network uses the differentiable Sinkhorn layer and annealing to get soft assignments of point correspondences from hybrid features learned from both spatial coordinates and local geometry. To further improve registration performance, we introduce a secondary network to predict optimal annealing parameters. Unlike some existing methods, our RPM-Net handles missing correspondences and point clouds with partial visibility. Experimental results show that our RPM-Net achieves state-of-the-art performance compared to existing non-deep learning and recent deep learning methods. Our source code is available at the project website https://github.com/yewzijian/RPMNet .

RPM-Net: Robust Point Matching using Learned Features

The presented paper introduces RPM-Net, a deep learning framework for robust rigid point cloud registration. RPM-Net builds upon the shortcomings of traditional approaches like Iterative Closest Point (ICP) by offering a framework less sensitive to initialization and capable of handling noisy and partially visible point clouds. It leverages a deep network to extract hybrid features combining spatial and geometric properties, thus enhancing robustness against noise and initialization variability.

Methodology

RPM-Net proposes an innovative approach using learned feature distances to improve initialization sensitivity. The traditional ICP algorithm iteratively alternates between determining point correspondences and calculating the rigid transformation. However, it falters in the presence of noise, outliers, and poor initializations, often getting trapped in local minima. RPM-Net addresses this by integrating a differentiable Sinkhorn layer with annealing to achieve soft point correspondences from learned hybrid features.

Feature Extraction

The feature extraction network in RPM-Net computes hybrid features from each point's spatial coordinates and local geometric properties, employing a PointNet architecture. This hybrid feature approach allows the model to incorporate comprehensive contextual information, thereby overcoming the limitations of purely spatial or geometric descriptors.

Annealing with Learned Parameters

Briefly deviating from traditional fixed schedules, RPM-Net utilizes a secondary neural network for predicting optimal annealing parameters, α and β. This predictive approach enables dynamic adjustment to the point cloud's current registration state, improving convergence behavior and registration precision.

Experimental Results

The paper demonstrates RPM-Net's efficacy through state-of-the-art performance on the ModelNet40 dataset across varied settings—clean, noisy, and partially visible data. RPM-Net exhibits significant improvements in alignment accuracy over prominent traditional methods like ICP and as well as recent deep learning solutions such as Deep Closest Point (DCP). For instance, under Gaussian noise settings, RPM-Net outperforms DCP-v2, achieving isotropic rotation and translation errors significantly lower than its predecessors.

Implications and Future Research

The integration of learned features and dynamic annealing parameters marks a substantial step forward in rigid point cloud registration, showcasing enhanced adaptability to diverse real-world scenarios. These innovations suggest potential expansions:

1. Generalization to Non-Rigid Registration: Extending RPM-Net's architecture to accommodate non-rigid transformations could benefit applications in dynamic environments, such as motion tracking.
2. Application to Larger Point Clouds: Efficient scaling mechanisms to manage high-dimensional point cloud data could broaden the method's applicability across different domains.
3. Integration with Sensor Technologies: Considering real-time data from multiple sensor modalities may further enhance robustness and accuracy, providing valuable insights into multi-sensor fusion applications.

The modular design, leveraging both spatial and learned characteristics, coupled with an adaptive annealing schedule, positions RPM-Net as a flexible and powerful tool in point cloud registration. This paper's contributions cement its role as a valuable asset in the computer vision and robotics communities, enabling more reliable and efficient registration outcomes.