Bolt3D
收藏资源简介:
Bolt3D数据集是由谷歌研究、牛津大学视觉几何组和谷歌DeepMind创建的一种大规模多视角一致性3D几何和外观数据集。该数据集通过应用最先进的密集3D重建技术,如MASt3R,到现有的多视角图像数据集上得到。它用于训练几何VAE和扩散模型,使得Bolt3D模型能够从一张或多张输入图像中直接生成3D场景表示,并在不到7秒内完成渲染。数据集的具体条数未在论文中提及。
The Bolt3D Dataset is a large-scale multi-view consistent 3D geometry and appearance dataset developed by Google Research, the Visual Geometry Group at the University of Oxford, and Google DeepMind. This dataset is derived by applying state-of-the-art dense 3D reconstruction techniques such as MASt3R to existing multi-view image datasets. It is used for training geometric VAEs and diffusion models, enabling the Bolt3D model to directly generate 3D scene representations from one or multiple input images and complete rendering within less than 7 seconds. The exact number of samples in this dataset is not mentioned in the accompanying paper.




