DeepGMR: Learning Latent Gaussian Mixture Models for Registration (2008.09088v1)

Abstract: Point cloud registration is a fundamental problem in 3D computer vision, graphics and robotics. For the last few decades, existing registration algorithms have struggled in situations with large transformations, noise, and time constraints. In this paper, we introduce Deep Gaussian Mixture Registration (DeepGMR), the first learning-based registration method that explicitly leverages a probabilistic registration paradigm by formulating registration as the minimization of KL-divergence between two probability distributions modeled as mixtures of Gaussians. We design a neural network that extracts pose-invariant correspondences between raw point clouds and Gaussian Mixture Model (GMM) parameters and two differentiable compute blocks that recover the optimal transformation from matched GMM parameters. This construction allows the network learn an SE(3)-invariant feature space, producing a global registration method that is real-time, generalizable, and robust to noise. Across synthetic and real-world data, our proposed method shows favorable performance when compared with state-of-the-art geometry-based and learning-based registration methods.

• The paper introduces a probabilistic framework for point cloud registration by modeling data with latent Gaussian mixtures to enhance noise robustness.
• The paper employs a neural network to learn SE(3)-invariant features, establishing pose-invariant correspondences without iterative procedures.
• The paper demonstrates efficient, real-time processing by computing optimal transformations in a single step, achieving high recall rates on benchmark datasets.

DeepGMR: Learning Latent Gaussian Mixture Models for Registration

The paper "DeepGMR: Learning Latent Gaussian Mixture Models for Registration" presents a novel approach to point cloud registration by introducing a method called Deep Gaussian Mixture Registration (DeepGMR). This approach is distinct in its explicit use of a probabilistic framework, articulated through a neural network architecture that models registration as the minimization of KL-divergence between two Gaussian Mixture Models (GMMs) of point clouds.

Technical Contributions

1. Probabilistic Registration Paradigm: DeepGMR utilizes probability distributions to model point clouds, contrasting with traditional methods focusing purely on geometric matching. By leveraging GMMs, the method captures and manipulates distributional properties, offering robustness against noise and arbitrary initial transformations.
2. Neural Network for Pose-Invariant Correspondences: The core of DeepGMR is a neural network trained to establish pose-invariant correspondences between points and GMM parameters. This network learns SE(3)-invariant features, contributing to a global registration method that is both generalizable and computationally efficient.
3. Elimination of Iterative Procedures: Conventional GMM-based methods typically rely on iterative Expectation Maximization (EM) algorithms, which can be computationally expensive. DeepGMR avoids iterative optimization by solving for optimal transformations in a single step utilizing learned point-to-GMM correspondences.

Numerical Results and Benchmarks

DeepGMR demonstrates favorable performance across synthetic datasets from ModelNet40 and real-world datasets such as ICL-NUIM. For instance, on noisy and real-world datasets, DeepGMR outperforms existing registration methods like ICP, FGR, and PointNetLK, achieving high recall rates with significantly reduced computational costs. The method processes up to 50 frames per second, indicating its suitability for real-time applications.

Implications and Future Directions

The introduction of DeepGMR marks a departure from geometry-centric methods to probabilistic models enriched by learning-based strategies. This shift enables handling larger transformations and noise robustly, with reduced sensitivity to initial conditions. Moreover, the integration of differentiable layers within a non-iterative framework promises potential applications where real-time performance is critical, such as robotics and autonomous navigation.

Future research could look into enhancing DeepGMR's capability to handle partial observations or occlusions, a common challenge in practical scenarios involving dynamic environments. Additionally, further exploration into combining probabilistic registration with other forms of deep learning can yield more robust feature representations, potentially broadening the application to varied domains, including medical imaging and augmented reality.

Overall, the presented work showcases a significant advancement in point cloud registration, leveraging the strengths of probabilistic modeling and neural networks to achieve fast, accurate, and robust registration under diverse and challenging conditions.