TL;DR
This paper introduces a novel framework for LiDAR point cloud super-resolution that treats sparse 3D data as structured 2D range images. By combining local convolutional filters with global State-Space Models (Mamba) and applying vertical spatial priors, the authors achieve state-of-the-art reconstruction quality, effectively eliminating the "over-smoothing" and "spike artifacts" that plague traditional methods.
What Problem Does This Paper Address?
LiDAR sensors are essential for autonomous driving, but they have a physical limitation: point density drops significantly as distance increases. This leaves distant objects sparse and blurry, hindering precise perception.
To fix this, researchers often use range-image-based super-resolution, projecting 3D points onto a 2D grid (like a panoramic photo where each pixel is a distance value). However, this approach faces three major hurdles:
- Vertical Nonuniformity: The top of a range image (distant sky/background) is much sparser and noisier than the bottom (nearby road), yet standard models treat them the same.
- The Smoothing Dilemma: Purely convolutional networks have small "receptive fields" and miss global context, while global attention models often over-smooth sharp edges, like the crisp outline of a car.
- Back-Projection Spikes: When converting a smoothed 2D image back into 3D space, tiny interpolation errors at object boundaries turn into giant "spikes" or floating artifacts.
Methods
The authors propose a three-pronged architecture to tackle these issues:
- LG-SSM Block (Local-Global Joint Modeling): This is the core engine. It uses depthwise separable convolutions to capture sharp local edges and textures, then passes the features to a Selective State-Space Model (SSM)—specifically inspired by the Mamba architecture. The SSM establishes long-range dependencies across the entire panoramic image with linear complexity, balancing fine detail with global consistency.
- SpatialGateFuse: Instead of simple skip-connections, this module uses a "vertical spatial prior." It applies a gated mask that recognizes the upper part of the image is less reliable. It adaptively enhances features in these sparse regions to prevent noise from polluting the decoder.
- 2D Median-Residual Denoising: A clever post-processing step. Before the final 3D back-projection, the model compares predicted depths against a local median. If a pixel deviates too much (a sign of an interpolation artifact), it is filtered out, preventing 3D "spikes."
Datasets
The model was trained and evaluated on two industry-standard autonomous driving datasets:
- KITTI-360: The authors downsampled 64-beam LiDAR scans to 16 beams for input, tasking the model with reconstructing the original 64-beam density.
- nuScenes: A more challenging setup where 8-beam inputs were upsampled to a 32-beam target resolution.
Key Insights and Ideas
- Mamba for LiDAR: While Transformers are popular for global modeling, their quadratic complexity is a burden for high-resolution range images. The use of SSMs (Mamba) allows for global context with the efficiency of a CNN.
- Geometry-Aware Fusion: The "aha moment" here is realizing that a range image isn't just a picture; it's a projection of a physical scan. By baking the vertical scanning geometry (the "vertical prior") into the fusion process, the model learns to be skeptical of noisy, distant data.
- Denoising in 2D for 3D Gains: Most denoising happens in 3D space, which is computationally expensive. This paper proves that identifying "anomalous residuals" in the 2D image domain is a highly effective and cheaper way to prevent 3D artifacts.
Key Results
The proposed method outperformed several recent baselines, including TULIP, SRFormer, and SRMamba, across both 2D and 3D metrics.
- KITTI-360: Achieved an IoU (Intersection over Union) of 0.4703, a significant jump over SRMamba (0.4389). It also halved the Chamfer Distance (CD) compared to most competitors, indicating much higher 3D structural accuracy.
- nuScenes: Reached an IoU of 0.3593, maintaining a lead even in very sparse (8-beam) scenarios.
- Qualitative Success: Visual results show that the method preserves the gaps between scan lines on the road and the sharp silhouettes of vehicles, whereas other methods often "melt" these structures together.
Limitations and Future Work
Not discussed in the paper. However, typical challenges for this type of work include the computational overhead of the SSM blocks compared to simple CNNs and the potential sensitivity to specific LiDAR calibration parameters when projecting range images.
Summary updated