Robust Point Cloud Registration Framework Based on Deep Graph Matching (2103.04256v1)

Abstract: 3D point cloud registration is a fundamental problem in computer vision and robotics. There has been extensive research in this area, but existing methods meet great challenges in situations with a large proportion of outliers and time constraints, but without good transformation initialization. Recently, a series of learning-based algorithms have been introduced and show advantages in speed. Many of them are based on correspondences between the two point clouds, so they do not rely on transformation initialization. However, these learning-based methods are sensitive to outliers, which lead to more incorrect correspondences. In this paper, we propose a novel deep graph matchingbased framework for point cloud registration. Specifically, we first transform point clouds into graphs and extract deep features for each point. Then, we develop a module based on deep graph matching to calculate a soft correspondence matrix. By using graph matching, not only the local geometry of each point but also its structure and topology in a larger range are considered in establishing correspondences, so that more correct correspondences are found. We train the network with a loss directly defined on the correspondences, and in the test stage the soft correspondences are transformed into hard one-to-one correspondences so that registration can be performed by singular value decomposition. Furthermore, we introduce a transformer-based method to generate edges for graph construction, which further improves the quality of the correspondences. Extensive experiments on registering clean, noisy, partial-to-partial and unseen category point clouds show that the proposed method achieves state-of-the-art performance. The code will be made publicly available at https://github.com/fukexue/RGM.

• The paper presents a novel deep graph matching framework that integrates local geometry and global structure for accurate 3D point cloud alignment.
• It employs an AIS module with soft correspondence refinement and SVD-based transformation estimation to robustly handle outliers.
• Extensive experiments on clean, noisy, and partial datasets validate its superior performance over traditional ICP and learning-based methods.

Robust Point Cloud Registration Framework Based on Deep Graph Matching

The paper "Robust Point Cloud Registration Framework Based on Deep Graph Matching" presents a novel approach to 3D point cloud registration, which is an essential task in computer vision and robotics. This task involves aligning two sets of 3D data points, optimizing for a rigid transformation composed of rotation and translation. While traditional methods such as the Iterative Closest Point (ICP) algorithm are computationally intensive and sensitive to initial transformations, this paper proposes leveraging deep learning methods to address the challenges posed by outliers and time constraints in point cloud registration.

Methodology

The core contribution of the paper is the introduction of a deep graph matching framework, named RGM, which effectively handles scenarios involving a substantial proportion of outliers. The approach stands out by transforming 3D point clouds into graphs, allowing both local geometry and broader structural and topological information to be incorporated into the matching process. This is a departure from many existing learning-based methods that rely solely on correspondences and are vulnerable to outliers, often resulting in incorrect matches.

The framework consists of several components:

1. Local Feature Extraction: Using a neural network, local features for each point are distilled from its geometric neighborhood, setting the groundwork for subsequent graph construction.
2. Graph Construction and Feature Extraction: Graphs are created with nodes representing points and edges generated through a transformer-based method. This step captures both local attributes and higher-order relationships.
3. AIS Module for Soft Correspondence Matrix Calculation: Employing deep graph matching techniques, an affinity matrix is computed, and intermediate soft correspondences are refined using instance normalization and the Sinkhorn algorithm. This process helps to filter out incorrect correspondences by considering a global context, complemented by iterative updates reminiscent of ICP methodology.
4. Transformation Estimation Using Singular Value Decomposition (SVD): Hard correspondences are derived from the soft matrix using a Linear Assignment Problem (LAP) solver, enabling the estimation of the optimal transformation via SVD.

Results and Evaluation

The efficacy of the proposed system is validated through extensive experimentation across varying complexities of the registration task, including clean, noisy, partial-to-partial, and previously unseen categorically distinct datasets. The proposed method consistently outperforms existing traditional and learning-based methods in terms of accuracy and robustness. Key metrics used in evaluation include rotational and translational errors alongside the Clip Chamfer Distance (CCD) which quantifies the registration accuracy by measuring the proximity of transformed point clouds.

Implications and Future Work

The introduction of deep graph matching within the context of point cloud registration offers a substantial advance in tackling the limitations of existing methodologies, particularly in handling discontinuities and partial overlaps in data. Furthermore, the AIS module showcases the potential for using graph-based structures to represent complex data relationships effectively.

Practically, this framework has significant implications for applications requiring real-time processing and robustness against data imperfections, such as autonomous driving and surgical navigation. The paper suggests that the integration of structural graph information with local geometric features could be further developed, potentially extending to other complex registration tasks like 2D-3D alignment and dynamic scene understanding.

In conclusion, the proposed framework not only achieves state-of-the-art results but also establishes a foundation for the application of graph matching techniques in new domains within computer vision and robotics. Future developments may explore various graph topologies or incorporate additional data modalities to enhance the adaptability and scalability of such models.