Synth4D

为了实现与SemanticKITTI和nuScenes的完全兼容，我们展示了使用CARLA模拟器创建的Synth4D。Tab.1将Synth4D与其他合成数据集进行了比较。Synth4D由两组点云序列组成，一组与SemanticKITTI (Velodyne HDL64E) 兼容，另一组与nuScenes (Velodyne HDL32E) 兼容。每组由20k标记的点云组成。Synth4D是使用车辆在四种场景 (城镇，公路，农村地区和城市)。由于UDA要求源和目标之间的标签一致，因此我们使用给注释器的原始说明将Synth4D的标签与SemanticKITTI/nuScenes的标签进行了映射，从而产生了八个宏类: 车辆，行人，道路，人行道，地形，人造，植被和未标记。To achieve full compatibility with SemanticKITTI and nuScenes, we present Synth4D—a synthetic dataset created using the CARLA simulator. Table 1 compares Synth4D with other existing synthetic datasets. Synth4D consists of two sets of point cloud sequences: one compatible with SemanticKITTI (Velodyne HDL64E) and the other compatible with nuScenes (Velodyne HDL32E). Each set comprises 20k annotated point clouds. Synth4D is generated with vehicles operating in four scenarios: town, highway, rural area, and urban district. Since unsupervised domain adaptation (UDA) requires consistent labels between source and target domains, we mapped the labels of Synth4D to those of SemanticKITTI/nuScenes using the original instructions provided to annotators, resulting in eight macro-classes: vehicle, pedestrian, road, sidewalk, terrain, man-made, vegetation, and unlabeled.

搜集汇总

背景与挑战

背景概述

Synth4D是一个用于自动驾驶研究的合成点云数据集，由CARLA模拟器生成，包含120k个标记点云，旨在与SemanticKITTI和nuScenes完全兼容。数据集分为两组序列，分别匹配不同激光雷达配置，覆盖城镇、公路、农村和城市四种场景，标签映射为八个宏类，如车辆、行人和道路等。它由东京大学等机构于2022年发布，适用于领域自适应和3D语义分割任务。