Robust Multiview Point Cloud Registration with Reliable Pose Graph Initialization and History Reweighting (2304.00467v1)

Abstract: In this paper, we present a new method for the multiview registration of point cloud. Previous multiview registration methods rely on exhaustive pairwise registration to construct a densely-connected pose graph and apply Iteratively Reweighted Least Square (IRLS) on the pose graph to compute the scan poses. However, constructing a densely-connected graph is time-consuming and contains lots of outlier edges, which makes the subsequent IRLS struggle to find correct poses. To address the above problems, we first propose to use a neural network to estimate the overlap between scan pairs, which enables us to construct a sparse but reliable pose graph. Then, we design a novel history reweighting function in the IRLS scheme, which has strong robustness to outlier edges on the graph. In comparison with existing multiview registration methods, our method achieves 11% higher registration recall on the 3DMatch dataset and ~13% lower registration errors on the ScanNet dataset while reducing ~70% required pairwise registrations. Comprehensive ablation studies are conducted to demonstrate the effectiveness of our designs.

• The paper introduces a novel neural network-based sparse pose graph initialization and history reweighting technique that achieves an 11% increase in recall and a 13% reduction in registration errors.
• It leverages global feature extraction and a closed-form iterative optimization to synchronize rotations and translations efficiently across multiple point clouds.
• The approach reduces computational complexity from quadratic to linear scale by requiring approximately 70% fewer pairwise registrations, benefitting 3D reconstruction and autonomous navigation.

Overview of Robust Multiview Point Cloud Registration

The paper "Robust Multiview Point Cloud Registration with Reliable Pose Graph Initialization and History Reweighting" introduces a novel approach to multiview registration of point clouds by leveraging both a neural network for graph initialization and a refined optimization method. The authors aim to address the inefficiencies and inaccuracies present in traditional methods, particularly those that rely on exhaustive pairwise registration for constructing dense pose graphs prone to outliers.

Key Contributions

1. Sparse Pose Graph Initialization: The authors propose constructing a sparse and reliable pose graph using a neural network to estimate overlap between scan pairs. This method significantly reduces the required number of pairwise registrations, minimizing computational complexity from $O(N^2)$ to $O(N)$.
2. History Reweighting in IRLS: A novel history reweighting function within the Iteratively Reweighted Least Squares (IRLS) framework improves robustness to outliers. This approach leads to more accurate registration results when compared to traditional IRLS implementations that often get trapped in local minima due to outliers.
3. Substantial Performance Improvements: The method is tested on three benchmark datasets—3DMatch, ScanNet, and ETH. The proposed approach yields an 11% increase in registration recall on the 3DMatch dataset and a 13% reduction in registration errors on the ScanNet dataset compared to existing techniques, while using approximately 70% fewer pairwise registrations.

Methodological Insights

• Global Feature Extraction: For each point cloud, a global feature vector is computed through a deep neural network, which aids in estimating pairwise overlap scores efficiently.
• Iterative Optimization: The method utilizes a closed-form solution to synchronize rotations and translations across the multiview data. The approach's resilience to outlier influence is fortified through a history-reweighting mechanism that aggregates residuals from previous iterations to influence current weight adjustments, similar to momentum-based optimization strategies.

Practical and Theoretical Implications

This paper presents significant advancements in efficiency and accuracy for multiview point cloud registration, which are vital for applications such as 3D reconstruction and autonomous navigation. The usage of a neural network for graph initialization introduces a new dimension to registration techniques by combining learning-based predictions with traditional optimization, suggesting potential future developments where hybrid models could increasingly replace purely deterministic approaches.

Future Developments

Potential future work may involve enhancing the generalization capabilities of the neural network when applied to unseen environments or varying conditions, as well as integrating this approach with other deep learning paradigms to holistically address end-to-end 3D scene understanding tasks. There is also room for exploring the application of this methodology in real-time systems, where speedy and reliable registration is critical.

The research opens avenues not only for further efficiency enhancements in registration algorithms but also offers a framework where neural network-driven solutions can directly augment or replace components of classical geometric methods in computer vision tasks.