BlendedMVS 多视图立体匹配数据集

BlendedMVS 是一个大规模的多视图立体匹配数据集，为基于学习的 MVS (Multi-view Stereo) 算法提供了充足的训练 ground truth 。该数据集是通过应用 3D 重建 pipeline，从精选场景图像中恢复高质量纹理 mesh 而生成的，包含 17,000 多个高分辨率图像，涵盖了各种场景，包括城市、建筑、雕塑和小型物体。实验表明：和其他数据集相比，使用 BlendedMVS 训练的网络模型具有更好的泛化能力。

BlendedMVS is a large-scale multi-view stereo (MVS) dataset that provides sufficient training ground truth for learning-based MVS algorithms. This dataset is generated by applying a 3D reconstruction pipeline to recover high-quality textured meshes from carefully curated scene images, containing over 17,000 high-resolution images covering various scenarios including urban scenes, buildings, sculptures and small objects. Experimental results demonstrate that network models trained with BlendedMVS exhibit better generalization performance compared to those trained on other datasets.

搜集汇总

背景与挑战

背景概述

BlendedMVS是一个大规模多视图立体匹配数据集，包含17,000多张高分辨率图像，涵盖城市、建筑等多种场景，通过3D重建技术提供高质量ground truth数据，能有效提升网络模型的泛化能力。