WebVi3D
收藏资源简介:
WebVi3D是由北京人工智能研究院创建的一个大规模多视角图像数据集,包含从1600万视频片段中提取的3.2亿帧图像。该数据集通过自动化的数据筛选流程,过滤掉动态内容和视角变化不足的视频,确保了数据的高质量和多样性。数据集的创建过程包括视频的时空下采样、动态场景识别、非刚性动态过滤和视角变化跟踪等步骤。WebVi3D主要应用于3D内容生成领域,旨在解决现有3D数据集规模有限、成本高昂的问题,支持从稀疏视图到3D生成、3D编辑等多种任务。
WebVi3D is a large-scale multi-view image dataset developed by the Beijing Institute of Artificial Intelligence. It contains 320 million frames of images extracted from 16 million video clips. Through an automated data filtering pipeline, this dataset filters out videos with dynamic content and insufficient viewpoint changes, ensuring high data quality and diversity. The dataset construction process includes steps such as spatiotemporal downsampling of videos, dynamic scene recognition, non-rigid dynamic filtering, and viewpoint change tracking. WebVi3D is primarily applied in the field of 3D content generation, aiming to address the limited scale and high cost issues of existing 3D datasets, and supports various tasks such as sparse view-to-3D generation and 3D editing.




