UniStereo
收藏资源简介:
UniStereo是由香港科技大学与快手科技联合构建的首个大规模统一立体视频转换数据集,涵盖平行和汇聚两种立体格式。该数据集包含约10.3万条高质量立体视频对,其中Stereo4D子集提供10万条平行格式数据,3DMovie子集贡献4.8万条电影级汇聚格式数据,所有视频均经标准化处理至832×480分辨率及16帧率,并配备自动化生成的文本描述。数据集通过专业预处理流程整合VR180视频与3D电影源,解决了立体视频领域格式割裂与评估标准不统一的问题,为单目到立体视频转换模型的训练与公平评测提供重要基础。
UniStereo is the first large-scale unified stereoscopic video conversion dataset jointly constructed by The Hong Kong University of Science and Technology and Kuaishou Technology, covering two stereoscopic formats: parallel and convergent. This dataset contains approximately 103,000 high-quality stereoscopic video pairs, where the Stereo4D subset provides 100,000 parallel-format samples, and the 3DMovie subset contributes 48,000 movie-level convergent-format samples. All videos have been standardized to a resolution of 832×480 and a frame rate of 16 fps, and are accompanied by automatically generated textual descriptions. Through a professional preprocessing pipeline, the dataset integrates VR180 videos and 3D movie sources, addressing the issues of format fragmentation and inconsistent evaluation standards in the stereoscopic video field. It provides an important foundation for the training and fair evaluation of monocular-to-stereoscopic video conversion models.
StereoPilot数据集概述
数据集名称
UniStereo
数据集简介
UniStereo是首个大规模的统一立体视频数据集,同时包含平行和汇聚两种立体格式。
数据集构成
UniStereo由两部分组成:
- 3DMovie:包含来自3D电影的汇聚立体格式数据。
- Stereo4D:包含平行立体格式数据(即将发布)。
数据处理
详细的数据处理说明请参考:https://github.com/KlingTeam/StereoPilot/blob/main/StereoPilot_Dataprocess/
相关研究
该数据集为论文《StereoPilot: Learning Unified and Efficient Stereo Conversion via Generative Priors》的研究组成部分,旨在支持高效、统一的多格式立体转换模型的训练与评估。




