BlendedMVS is a large-scale multi-view stereo matching dataset that provides sufficient training ground truth for the learning-based MVS (Multi-view Stereo) algorithm. The dataset is generated by applying a 3D reconstruction pipeline to recover high-quality texture meshes from selected scene images. It contains more than 17,000 high-resolution images covering a variety of scenes, including cities, buildings, sculptures, and small objects. Experiments show that compared with other datasets, the network model trained using BlendedMVS has better generalization ability.