半合成数据集
收藏资源简介:
半合成数据集是由商汤科技研究院开发,用于立体匹配研究的大规模数据集。该数据集通过从真实场景中提取图像块并将其纹理应用于生成的几何形状上,快速合成大量接近真实场景的纹理数据,以缩小合成数据与真实数据之间的差距。数据集包含约35000对合成图像,具有精确的密集视差图和高场景多样性。创建过程中,使用Blender软件生成所需立体数据,包括左右图像和密集视差地面实况。该数据集主要应用于立体匹配领域,旨在解决深度学习模型在立体匹配任务中对大量训练数据的依赖问题,以及合成数据与真实数据之间的领域差异问题。
This semi-synthetic dataset is a large-scale dataset developed by SenseTime Research Institute for stereo matching research. It quickly synthesizes a large amount of texture data that closely resembles real-world scenes by extracting image patches from real scenes and applying their textures to generated geometric shapes, thereby narrowing the domain gap between synthetic and real-world data. The dataset contains approximately 35,000 pairs of synthetic images, equipped with accurate dense disparity maps and high scene diversity. During its creation, Blender software was used to generate the required stereo data, including left-right image pairs and dense disparity ground truth. This dataset is primarily applied in the field of stereo matching, aiming to address two core issues: the heavy dependence of deep learning models on large-scale training data for stereo matching tasks, and the domain gap between synthetic and real-world data.




