Slice Transformer and Self-supervised Learning for 6DoF Localization in 3D Point Cloud Maps

Muhammad Ibrahim, Naveed Akhtar, Saeed Anwar, Michael J. Wise, Ajmal Mian

Abstract

Precise 6-degree-of-freedom (6DoF) localization is a critical requirement for autonomous vehicle navigation, particularly in "urban canyons" where Global Navigation Satellite Systems (GNSS) often fail. While 3D LiDAR-based localization offers a robust alternative, traditional methods are computationally intensive. This paper introduces the Slice Transformer, a novel architecture that leverages multi-head attention for outdoor LiDAR localization. By combining a unique "slicing" tokenization strategy with a self-supervised quadrant-shuffling pretext task, the proposed method achieves state-of-the-art precision. We also introduce the Perth-WA dataset, a large-scale 3D LiDAR benchmark, to validate the model's performance in complex urban environments.

Introduction

For autonomous systems, localization is a fundamental safety concern. In dense urban environments—characterized by tall buildings, tunnels, and bridges—satellite signals are frequently obstructed or reflected, leading to significant positioning errors. 3D LiDAR-based localization has emerged as the gold standard for these scenarios. However, the field is shifting from traditional point cloud registration toward "map compression" techniques, where deep learning models learn to regress a vehicle's pose directly from a single LiDAR scan.

Despite their success in other domains, standard deep learning architectures like Convolutional Neural Networks (CNNs) and Multi-Layer Perceptrons (MLPs) struggle with the unstructured and sparse nature of outdoor point clouds. Transformers offer a promising alternative due to their ability to model complex global relationships, but they typically require massive amounts of labeled data—a resource that is notoriously difficult to acquire for 3D outdoor environments. This work addresses these challenges through architectural innovation and self-supervised learning.

Methods

The Slice Transformer Architecture

The core innovation of this work is the Slice Transformer. Unlike traditional models that attempt to process a 3D scan as an unstructured aggregate, this architecture exploits the axial symmetry inherent in spinning LiDAR sensors. The 360° LiDAR frame is divided into 36 overlapping "slices," which serve as the input tokens for the Transformer.

By treating these slices as sequential data, the model utilizes Multi-Head Attention (MHA) to identify and weigh the importance of specific environmental features, such as building corners or unique structural poles. This approach allows the model to focus on the most distinct landmarks within a scene, improving the robustness of the resulting pose estimation.

Self-Supervised Pre-text Task

To mitigate the data-hungry nature of Transformers, we implemented a self-supervised "quadrant-shuffling" pretext task. In this phase, a LiDAR frame is divided into four quadrants and shuffled. The model is then tasked with predicting the correct permutation out of 24 possible arrangements.

This pre-training phase utilized 250,000 unlabeled frames, allowing the model to learn the intrinsic geometric properties of urban environments without requiring human-annotated ground truth. For the final localization task, the classification head used for shuffling is replaced with a Regression Module designed to predict 6DoF pose: translation (X, Y, Z) and rotation (yaw, roll, pitch).

Datasets

The model was evaluated on several benchmarks, including a significant new dataset introduced in this study:

  • Perth-WA: A large-scale dataset covering 4km² of the Perth Central Business District in Western Australia. It contains 30,000 frames with 6DoF annotations. Notably, the ground truth was established via LiDAR frame registration rather than IMU data to ensure higher precision.
  • Apollo-SouthBay: A standard large-scale dataset collected in San Francisco, used to benchmark against existing state-of-the-art methods.
  • ModelNet40 & ScanObjectNN: Standard benchmarks used to evaluate the versatility of the pre-trained backbone for general 3D object classification.

Results and Discussion

The Slice Transformer consistently outperformed existing localization methods, such as PointLoc, across multiple metrics. On the Perth-WA dataset, the pre-trained model achieved a mean translation error of 5.23 meters and a rotation error of 0.59°, a substantial improvement over the non-pre-trained baseline (13.25m and 1.00°, respectively).

Key insights from the experimental results include:

  • Effectiveness of Pre-training: The model achieved approximately 95% accuracy on the quadrant-shuffling task. This "warm start" was essential for achieving high localization accuracy, proving that self-supervised learning can effectively bridge the data gap for 3D Transformers.
  • Generalization: When tested on the ScanObjectNN classification benchmark, the model set a new state-of-the-art accuracy of 84.5%. This suggests that the features learned during the shuffling task are highly robust and transferable to other 3D vision tasks.
  • Urban Performance: On the Apollo-SouthBay dataset, the Slice Transformer achieved the lowest average errors across all tested routes, outperforming established methods by significant margins.

Conclusion

The Slice Transformer represents a successful application of multi-head attention to large-scale outdoor point clouds. By combining axial slicing with self-supervised pre-training, the model overcomes the traditional limitations of Transformers in the 3D domain. While the current implementation focuses on points within a 70-meter radius to maintain density, future work will explore more complex tokenization methods to handle sparse environments and dynamic changes in the map. The release of the Perth-WA dataset provides a new benchmark for the community to further refine 6DoF localization techniques.

Summary updated

BibTeX
@inproceedings{ibrahim2023slice,
  author = {Muhammad Ibrahim and Naveed Akhtar and Saeed Anwar and Michael J. Wise and Ajmal Mian},
  title = {{Slice Transformer and Self-supervised Learning for 6DoF Localization in 3D Point Cloud Maps}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
  url = {https://openreview.net/forum?id=vx904ZDgvR},
}
Figure