Abstract
Autonomous driving systems rely heavily on Lidar sensors, yet performance often degrades when models encounter sensors with different resolutions than their training data—a phenomenon known as domain shift. We present a lightweight, one-stage upsampling method that leverages Sliced Wasserstein Distance (SWD) and edge-aware convolutions to reconstruct high-fidelity scan patterns from low-resolution data. Our approach effectively mitigates domain shift, preserving the characteristic "striped" geometry of Lidar sweeps and significantly improving downstream object detection performance.
Introduction
In autonomous driving, Lidar sensors provide the critical spatial awareness required for safe navigation. However, not all sensors are created equal; a model trained on a high-end 64-line Lidar sensor often fails when deployed on a more affordable 16-line or 32-line sensor. To address this without the prohibitive cost of re-annotating thousands of hours of data, researchers utilize upsampling (super-resolution for point clouds).
Lidar data is notoriously difficult to upsample because it is sparse, irregular, and follows a very specific "striped" pattern based on the sensor's laser sweeps. Existing methods often produce noisy point clouds or fail to maintain these distinct scan lines, which are crucial for downstream tasks like object detection.
Proposed Architecture
The researchers introduce a streamlined, one-stage architecture designed for efficiency and geometric fidelity. Unlike traditional "coarse-to-fine" networks, this model achieves high-quality reconstruction in a single pass through three primary stages:
- Feature Extraction (Ext): Combines 1D convolutions with EdgeConv layers. EdgeConv utilizes a k-Nearest Neighbors (kNN) approach to find local clusters, allowing the network to learn geometric features like edges and surfaces rather than just individual point coordinates.
- Feature Expansion (Exp): Uses additional EdgeConv layers to expand the feature space, preparing the network to generate new, spatially coherent points.
- Coordinate Reconstruction (Rec): A set of linear layers transforms the expanded features back into 3D coordinates.
Sliced Wasserstein Distance (SWD)
The core innovation of this work is the loss function. Standard metrics like Chamfer Distance (CD) suffer from "CD blindness"—they do not penalize uneven point distributions, often resulting in "clumpy" point clouds. While Earth Mover’s Distance (EMD) is more descriptive, it is computationally expensive and prone to distortion.
To solve this, the authors utilize Sliced Wasserstein Distance (SWD). SWD projects 3D point distributions onto random 1D lines. In 1D, the "optimal transport" problem has a simple closed-form solution: sorting. This makes SWD computationally efficient while remaining highly sensitive to the fine-grained geometric structures (the "lines") of a Lidar sweep.
Experimental Results
The model was evaluated on the SemanticKITTI and Waymo Open Datasets. Quantitative results demonstrate that the proposed method outperforms state-of-the-art models like PU-Net, 3PU, and PU-GAN in Hausdorff Distance (HD) and SWD metrics. For 2x upsampling on KITTI, it achieved an HD of 0.1694, compared to PU-GAN's 0.1892.
Qualitatively, the results are striking. While other methods produce scattered "clouds" of points, this method successfully reconstructs the missing horizontal scan lines, making the upsampled data look nearly identical to the high-resolution ground truth. Furthermore, testing the upsampled data with a PointPillars object detector improved the Mean Average Precision (mAP) for car detection from 59.7% (low-res) to 76.2%, effectively narrowing the gap toward the high-res baseline of 86.6%.
Conclusion
By combining edge-aware convolutions with the Sliced Wasserstein Distance, this work provides a robust solution for Lidar domain shift. The method preserves the essential "striped" geometry of Lidar scans, leading to substantial improvements in both visual fidelity and downstream perception accuracy. Future work will focus on scaling this architecture to handle full-frame Lidar scans in real-time.
Summary updated