ResLPR: A LiDAR Data Restoration Network and Benchmark for Robust Place Recognition Against Weather Corruptions

TL;DR

This paper addresses the vulnerability of LiDAR-based place recognition (LPR) to adverse weather by introducing ResLPRNet, a wavelet-based restoration network with a ContextGuide module that preprocesses weather-corrupted scans before LPR. It also presents ResLPR, a Weather-induced LPR benchmark comprising WeatherKITTI and WeatherNCLT, along with a mean stability rate metric to quantify robustness. Empirical results show that applying ResLPRNet significantly improves multiple pretrained LPR models across snow, fog, and rain, under short-term and long-term scenarios, while maintaining real-time performance. The work advances practical LPR reliability in autonomous driving by coupling restoration with a dedicated robustness benchmark and open-sourcing code and data for further research.

Abstract

LiDAR-based place recognition (LPR) is a key component for autonomous driving, and its resilience to environmental corruption is critical for safety in high-stakes applications. While state-of-the-art (SOTA) LPR methods perform well in clean weather, they still struggle with weather-induced corruption commonly encountered in driving scenarios. To tackle this, we propose ResLPRNet, a novel LiDAR data restoration network that largely enhances LPR performance under adverse weather by restoring corrupted LiDAR scans using a wavelet transform-based network. ResLPRNet is efficient, lightweight and can be integrated plug-and-play with pretrained LPR models without substantial additional computational cost. Given the lack of LPR datasets under adverse weather, we introduce ResLPR, a novel benchmark that examines SOTA LPR methods under a wide range of LiDAR distortions induced by severe snow, fog, and rain conditions. Experiments on our proposed WeatherKITTI and WeatherNCLT datasets demonstrate the resilience and notable gains achieved by using our restoration method with multiple LPR approaches in challenging weather scenarios. Our code and benchmark are publicly available here: https://github.com/nubot-nudt/ResLPR.

Figures (6)

• Figure 1: Our method restores LiDAR data under weather corruptions, improving the performance of a pretrained LPR model.
• Figure 2: Pipeline of ResLPRNet. It adopts a hierarchical encoder-decoder architecture to restore clean LiDAR data from noisy weather-induced inputs. The three-layer encoder and decoder employ similar WaveTransformer blocks, with the encoder performing wavelet decomposition and the decoder executing wavelet reconstruction. An additional Transformer block is applied to the bottleneck feature to enhance its spatial correlation. ContextGuide blocks in each decoder layer enable the network to adapt to different degradation.
• Figure 3: Details of the Feature Mixing Block, which includes two operations: spatial mixing and channel mixing. BN denotes batch normalization, GConv represents grouped convolution and FC is fully connected layer. Other symbol definitions are the same as in Fig. \ref{['restore_ov']}.
• Figure 4: Recall@N curves on the WeatherKITTI and WeatherNCLT datasets. (a) Snow results (b) Fog results (c) Rain results
• Figure 5: Qualitative visualizations of some corruption query examples, along with their top-1 retrieved matches on the WeatherKITTI and WeatherNCLT datasets using CVTNet. The red points in the point cloud signify the noise points, and the blue points denote the lost points. red boxes indicate incorrect retrieval results, while green boxes denote correct retrievals.