DeepLoc

DeepLoc 是一个大规模的城市户外定位数据集。该数据集目前由一个跨越 110 x 130 m 区域的场景组成，机器人以不同的驾驶模式多次穿越。数据集创建者使用基于 LiDAR 的 SLAM 系统，具有亚厘米和亚度的精度来计算提供作为 groundtruth 的姿势标签。数据集中的姿势大约间隔 0.5 m，是其他重定位数据集的两倍密集。
此外，对于每张图像，数据集创建者为十个类别提供像素级语义分割注释：背景、天空、道路、人行道、草、植被、建筑、杆和栅栏、动态和虚空。数据集分为火车和测试拆分，火车集包括七个循环，具有交替的驾驶风格，共计 2737 张图像，而测试集包括三个循环，总共 1173 张图像。该数据集还包含全球 GPS/INS 数据和 LiDAR 测量值。
该数据集对于基于视觉的应用程序（例如全局定位、相机重定位、语义分割、视觉里程计和闭环检测）可能非常具有挑战性，因为它包含大量照明、天气变化、重复结构、反射和透明玻璃建筑物。

DeepLoc is a large-scale urban outdoor localization dataset. Currently, this dataset consists of a scene spanning an area of 110 × 130 meters, which has been traversed multiple times by robots under different driving modes. The dataset creators used a LiDAR-based SLAM system with sub-centimeter and sub-degree accuracy to calculate the pose labels provided as ground truth. The poses in the dataset are spaced approximately every 0.5 meters, which is twice as dense as other relocalization datasets.

Additionally, for each image, the dataset creators provide pixel-level semantic segmentation annotations for ten categories: background, sky, road, sidewalk, grass, vegetation, building, pole and fence, dynamic, and void. The dataset is split into training and test sets: the training set includes seven loops with alternating driving styles, totaling 2,737 images, while the test set contains three loops with a total of 1,173 images. This dataset also includes global GPS/INS data and LiDAR measurements.

This dataset poses significant challenges for vision-based applications such as global localization, camera relocalization, semantic segmentation, visual odometry, and loop closure detection, as it contains extensive variations in illumination, weather, repeated structures, reflections, and transparent glass buildings.

搜集汇总

背景与挑战

背景概述

DeepLoc是一个大规模城市户外定位数据集，覆盖110x130米区域，提供亚厘米级精度的姿态标签和十类像素级语义分割注释。数据集包含训练集2737张图像和测试集1173张图像，适用于视觉定位、语义分割等任务，并涵盖了光照、天气等多种环境变化。