Abstract
High-resolution LiDAR sensors are critical for autonomous navigation but remain prohibitively expensive. Super-resolution (SR) offers a computational path to high-density data, yet traditional methods struggle with arbitrary upscaling factors and the "blurring" of sharp object boundaries. We present the Implicit LiDAR Network (ILN), which achieves arbitrary-scale super-resolution by predicting non-linear interpolation weights instead of raw depth values. By utilizing a Transformer-based self-attention mechanism, ILN effectively preserves geometric edges and converges significantly faster than previous implicit neural representations. Our results on a large-scale synthetic benchmark demonstrate significant improvements in both 2D range image accuracy and 3D point cloud reconstruction.
Introduction
LiDAR sensors provide the spatial awareness necessary for autonomous vehicles. However, the trade-off between resolution, cost, and frame rate often limits the density of captured point clouds. Super-resolution techniques aim to upscale low-resolution LiDAR data, but they face two primary challenges: fixed scaling factors in traditional CNN architectures and the "ghosting" artifacts produced by linear interpolation in recent implicit representation models like LIIF.
ILN addresses these issues by reframing the super-resolution task. Rather than regressing a depth value for a query coordinate from scratch, the network learns to predict optimal weights for blending neighboring pixels. This approach constrains the output to the local data manifold, preserving the sharp discontinuities characteristic of LiDAR data.
Methodology: Predicting Weights, Not Values
The core innovation of ILN is the shift from value prediction to weight prediction. By predicting weights, the model is constrained to stay within the range of observed data, making the learning process more stable and preserving sharp transitions found at object edges.
Architecture Overview
- Feature Encoder: A standard CNN extracts deep feature maps from the low-resolution input range image.
- Local Feature Extraction: For any query point, the model identifies the four nearest neighbor pixels and extracts their features, augmented with relative distance embeddings.
- Feature Transformer: The neighbor features are processed through a self-attention module. This allows the network to determine the geometric relevance of each neighbor to the query point based on local geometry.
- Weight Prediction: The network outputs four weights ($w_1$ to $w_4$) via a Softmax layer. The final depth $\hat{r}$ is calculated as the weighted sum of the actual input depths: $\hat{r} = \sum w_t \cdot r_t$.
Experiments and Results
To evaluate ILN, the authors developed a large-scale synthetic benchmark using the CARLA simulator, encompassing approximately 24,000 waypoints across 8 different urban scenes. Testing was conducted on both in-distribution and out-of-distribution resolutions to test arbitrary scaling capabilities.
Performance Comparison
ILN significantly outperforms the previous state-of-the-art, LIIF, across all metrics. In 3D reconstruction (measured by IoU), ILN achieved 0.331 compared to LIIF’s 0.236 at $128 \times 2048$ resolution.
Key Findings
- Edge Preservation: As shown in the visual results, ILN avoids the "ghosting" effects seen in LIIF, maintaining the integrity of object boundaries in 3D space. Because the weights are non-linear, the network can effectively "pick" one neighbor over another when a query point sits on an edge.
- Faster Convergence: ILN reaches peak performance much faster than value-prediction networks. The authors attribute this to the fact that the network doesn't have to "re-learn" the input depths; it only learns how to combine them.
- Robustness to Scale: Even when upscaling to $256 \times 4096$ (a resolution higher than seen in training), ILN maintained high performance, proving its "limitless" upscaling capability is more robust than prior methods.
Conclusion
The Implicit LiDAR Network (ILN) provides a robust framework for arbitrary-scale LiDAR super-resolution. By leveraging Transformer-based weight prediction, it overcomes the blurring artifacts of previous implicit methods while offering faster training and superior 3D reconstruction. Future work will focus on adapting ILN to real-world sensor noise and evaluating its impact on downstream autonomous driving tasks like real-time motion planning.
Summary updated