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

Wenqing Kuang¹, Xiongwei Zhao², Yehui Shen¹, Congcong Wen³, Huimin Lu¹, Zongtan Zhou¹, Xieyuanli Chen¹

¹National University of Defense Technology  ²Harbin Institute of Technology  ³New York University Abu Dhabi

About

ResLPR is a benchmark for LiDAR-based place recognition under adverse weather. ResLPRNet, a lightweight network, restores corrupted LiDAR scans using wavelet transform. It helps pretrained models enhance the robustness of place recognition in bad weather in a plug-and-play manner.

1. The visualizations of WeatherKITTI and WeatherNCLT proposed by this benchmark are as follows:

Visualization of corrupted point clouds of varying severity in WeatherKITTI and WeatherNCLT. The 🔴 red points in the point cloud signify the noise points, and the 🔵 blue points denote the lost points.

2. The visualizations of the results of three different preprocessing algorithms (including ResLPRNet) are as follows:

Data Preparation

Our datasets are hosted by Baidu Netdisk. Download the dataset via this ResLPR datasets. Kindly refer to DATA_PREPARE.md for the details to prepare the ¹WeatherKITTI, ²WeatherNCLT.