Spot-the-shift
收藏资源简介:
该数据集包含两个配置:default和training_data,主要用于地面街景图像的处理和分析。default配置包含939个测试样本,每个样本包含两幅图像路径、配对ID、城市、距离、视角方向以及变化单元(包括类型、动作、位置和最终掩码路径)。training_data配置包含5000个训练样本,每样本包含两幅图像路径、句子描述、城市、图像ID、翻转标志、结构性和非结构性编辑数量以及点信息(包括步骤、中心点、标签和类型)。数据集适用于基于街景的图像差异描述、变化检测以及大型多模态模型的空间推理能力评估。数据来源于Mapillary街景序列,已自动模糊人脸和车牌以保护隐私。数据集存在地理偏差,某些地区未包含在当前版本中。
This dataset comprises two configurations: default and training_data, primarily designed for ground-level street view image processing and analysis. The default configuration includes 939 test samples, each containing two image paths, a pair ID, city, distance, viewing direction, and change units (including type, action, position, and final mask path). The training_data configuration contains 5000 training samples, each with two image paths, sentence descriptions, city, image ID, flip flag, number of structural and non-structural edits, and point information (including step, center point, label, and type). This dataset is suitable for street view-based image difference description, change detection, and spatial reasoning capability evaluation of large multimodal models. The data is sourced from Mapillary street view sequences, and human faces and license plates have been automatically blurred to protect privacy. The dataset has geographic bias, and some regions are not included in the current version.




