PhotoSynth (PS) dataset

PhotoSynth (PS) 数据集是由印度理工学院孟买分校创建，旨在改进局部图像描述符的学习，特别用于提升补丁匹配的性能。该数据集包含30个场景，总计约750万条匹配对应关系，每个场景平均包含200张RGB图像，分辨率在2000px×1350px至800px×550px之间。数据集创建过程中采用了结构从运动(SFM)技术，并提供了包括位置、尺度、旋转等信息的RGB补丁，以及相机的内外参数。PS数据集广泛应用于计算机视觉任务，如3D重建、结构从运动(SFM)、宽基线匹配等，旨在解决真实世界场景中由于视角、尺度、光照变化等因素导致的匹配难题。

The PhotoSynth (PS) dataset was developed by the Indian Institute of Technology Bombay to advance the learning of local image descriptors, with a primary focus on enhancing patch matching performance. This dataset consists of 30 scenes, totaling approximately 7.5 million matching correspondences. Each scene contains an average of 200 RGB images, with resolutions ranging from 2000px × 1350px down to 800px × 550px. Structure from Motion (SfM) technology was utilized during the dataset's construction phase. The dataset provides RGB patches with metadata including position, scale, rotation, and other relevant attributes, as well as the intrinsic and extrinsic parameters of the associated cameras. The PS dataset is widely applied in computer vision tasks such as 3D reconstruction, Structure from Motion (SfM), wide-baseline matching, and others, aiming to resolve matching challenges in real-world scenes arising from variations in viewpoint, scale, illumination, and other influencing factors.