Dynamic RealEstate10K (D-RE10K); D-RE10K-iPhone
收藏资源简介:
Dynamic RealEstate10K是由弗吉尼亚大学构建的大规模动态场景数据集,包含1.5万条真实室内视频序列,涵盖相机移动和物体运动的复杂场景。该数据集通过YouTube房地产导览和宠物互动视频筛选构建,包含人类、宠物等动态元素,填补了现有静态NVS数据集的空白。其子集D-RE10K-iPhone提供配对的瞬态/干净视图基准,支持稀疏视图下的瞬态感知NVS评估。该数据集旨在推动动态环境中自监督新视角合成技术发展,解决传统方法在动态内容处理时的多视角一致性问题。
Dynamic RealEstate10K is a large-scale dynamic scene dataset constructed by the University of Virginia, which contains 15,000 real indoor video sequences covering complex scenarios involving camera motions and object movements. This dataset is built by filtering YouTube real estate tours and pet interaction videos, and includes dynamic elements such as humans and pets, filling the gap left by existing static Neural View Synthesis (NVS) datasets. Its subset D-RE10K-iPhone provides paired transient/clean view benchmarks, supporting transient-aware NVS evaluation under sparse-view settings. This dataset aims to promote the development of self-supervised novel view synthesis technologies in dynamic environments, and solve the multi-view consistency problem that traditional methods face when processing dynamic content.

- 1WildRayZer: Self-supervised Large View Synthesis in Dynamic Environments弗吉尼亚大学 · 2026年



