Abstract
Place recognition is a fundamental capability for autonomous vehicles and mobile robots, enabling them to identify previously visited locations. While LiDAR-based 3D point clouds offer stability against lighting and weather changes, training deep learning models for this task requires vast amounts of labeled data. We propose vLPD-Net (virtual Large-scale Point cloud Descriptor Network), a novel framework that bridges the gap between synthetic and real-world data. By combining a structure-aware registration network (S-ARN) with adversarial domain adaptation, vLPD-Net learns robust geometric descriptors. Our experiments demonstrate that training on synthetic data from Grand Theft Auto V and adapting to real-world environments yields performance that surpasses current state-of-the-art baselines on the Oxford RobotCar dataset.
Introduction
Reliable place recognition is essential for loop closure in SLAM (Simultaneous Localization and Mapping) systems. LiDAR sensors provide 3D structural information that is inherently more robust than visual data in challenging environments. However, the high cost of manual labeling for large-scale outdoor point clouds limits the availability of training data. Synthetic data from high-fidelity simulations offers a potential solution, but the "domain gap"—differences in noise, density, and sensor patterns between virtual and real worlds—often degrades model performance.
vLPD-Net addresses these challenges by treating point cloud registration as a "teacher" task to extract deeper structural knowledge and employing adversarial training to align virtual and real feature distributions.
Methodology
The vLPD-Net architecture is designed around three core components:
1. Feature Extraction
The network utilizes an enhanced LPD-Net backbone to extract multi-scale local features. It employs graph-based aggregation in Cartesian space to capture the local geometric context of each point, forming the basis for both local and global descriptors.
2. Structure-Aware Registration Network (S-ARN)
A key innovation of vLPD-Net is the S-ARN, which forces the network to learn features capable of aligning two disparate point clouds. Successful registration implies a deep understanding of geometric structure, which translates to more robust place recognition descriptors.
- Light-weighted Attention Mechanism (LAM): To handle large-scale outdoor scenes efficiently, vLPD-Net uses a simplified attention mechanism with $O(N)$ complexity, significantly faster than standard Transformers.
- Efficient Point Cloud Outlier Removal (EPCOR): This module filters non-matching points, ensuring that the registration process focuses on stable, shared structural features.
3. Adversarial Domain Adaptation
To mitigate the domain gap, a discriminator network is introduced. During training, the feature extractor attempts to generate domain-invariant features that "fool" the discriminator into being unable to distinguish between virtual (GTA-V) and real (Oxford RobotCar) data.
Experimental Results
The authors evaluated vLPD-Net using a custom synthetic dataset generated from Grand Theft Auto V (2.5km route, 1,016 frames) and the real-world Oxford RobotCar benchmark.
Place Recognition Performance
vLPD-Net demonstrated superior performance across multiple metrics:
- Real-to-Real: When trained and tested on real data, vLPD-Net achieved a 96.37% Recall@1%, outperforming the original LPD-Net (94.92%) and PointNetVLAD (81.01%).
- Virtual-to-Real: When trained on virtual data and adapted to the real world, vLPD-Net reached 73.14% Recall@1%, a significant improvement over the 62.61% achieved by LPD-Net without domain adaptation.
Registration Accuracy
On virtual datasets, vLPD-Net significantly outperformed traditional methods (ICP, FGR) and deep learning baselines (DCP) in both rotation and translation accuracy, validating the effectiveness of the S-ARN module.
Conclusion
vLPD-Net successfully demonstrates that synthetic data can be a powerful resource for 3D place recognition when paired with effective domain adaptation and structure-aware learning. By using registration as a proxy task, the network learns geometric descriptors that are robust to the noise and sparsity of real-world LiDAR data. Future work will explore scaling the synthetic training data and integrating the network into real-time SLAM pipelines for field testing.
Summary updated