TransLiDAR: A Dataset and Benchmark for Cross-Sensor Point Cloud Translation

Abstract

Autonomous vehicles are typically equipped with one primary and several auxiliary LiDAR sensors to generate point clouds of the environment. However, differences in structural design, resolution, and scanning mechanisms among LiDAR types lead to significant modality gaps, which hinder cross-sensor adaptation and joint learning. To address this challenge, we aim to enable effective point cloud translation across heterogeneous LiDAR sensors. To this end, we construct a cross-sensor point cloud translation dataset, TransLiDAR, containing paired point clouds from mechanical and hybrid semi-solid-state LiDARs captured in the same scenes. Based on this dataset, we further investigate point cloud translation methods and propose a framework called TransLiDAR-Net. Specifically, point clouds are projected into 2D range images, and a dual-stream network is used to separately extract foreground and background features. A feature interaction module is then introduced to facilitate information exchange between the two feature streams. Experimental results demonstrate that TransLiDAR-Net effectively translates point clouds between mechanical and hybrid semi-solid-state LiDARs, providing a practical solution and strong baseline for cross-sensor adaptation and joint learning.