Mapillary Street-level Sequences

Mapillary Street-Level Sequences (MSLS) is the largest, most diverse dataset for place recognition, containing 1.6 million images in a large number of short sequences. Spanning 30 cities on six continents, the dataset covers different seasons, weather and daylight conditions, various camera types and viewpoints, diverse architectural and structural settings (such as roadworks), and different levels of dynamic objects present in the scenes (such as moving pedestrians or cars). Each image comes with metadata and attributes relevant for further research: raw GPS coordinates, capture time, and compass angle, as well as attributes for day/night, and view direction (front-, back-, or side-facing). We have also run extensive benchmarks on our dataset with previous state-of-the-art methods for place recognition. The results show that training on MSLS improves performance due to the diversity of the dataset in geographical distribution, seasonal and temporal changes, and particularly day/night changes. Thanks to its wide geographical reach, diversity in scene characteristics, and sufficient size for training neural networks with large capacity, MSLS is the best dataset for pushing the state of the art in visual place recognition and its applications in practical settings across the world.

马皮拉利街景序列数据集（Mapillary Street-Level Sequences, MSLS）是当前规模最大、多样性最丰富的地点识别数据集，涵盖按大量短序列组织的共计160万张图像。该数据集覆盖六大洲的30座城市，包含不同季节、天气与光照条件，多样的相机类型与拍摄视角，丰富的建筑与场景结构（如道路施工场景），以及场景中各类动态物体（如移动的行人或车辆）。每张图像均附带支持后续研究的元数据与属性信息：原始GPS坐标、拍摄时刻、罗盘角度，以及昼夜属性与拍摄朝向（正向、背向或侧向）。我们还基于该数据集，使用此前地点识别领域的顶尖方法开展了全面的基准测试。测试结果表明，得益于该数据集在地理分布、季节与时序变化，尤其是昼夜变化上的多样性，在MSLS上进行训练可有效提升模型性能。凭借广泛的地理覆盖范围、丰富的场景特征多样性，以及足以支撑大容量神经网络训练的充足规模，MSLS是推动视觉地点识别技术及其全球实际应用领域前沿发展的最优数据集。

搜集汇总

背景与挑战

背景概述

Mapillary Street-level Sequences是一个大规模、多样化的地点识别数据集，包含160万张图像，覆盖30个城市和六大洲的不同季节、天气和光照条件。每张图像都附有GPS坐标、拍摄时间和罗盘角度等元数据，适用于视觉地点识别的研究和应用。