MIVE Dataset
收藏资源简介:
MIVE Dataset是由韩国科学技术院创建的一个多实例视频编辑数据集,旨在支持多实例视频编辑任务的评估。该数据集包含200个视频,涵盖了多种场景和实例,每个视频都配有实例级别的标注和掩码。数据集的创建过程结合了多实例采样和实例中心概率重分配技术,确保编辑的精确性和忠实性。该数据集主要应用于多实例视频编辑领域,旨在解决现有方法在多对象编辑中的编辑泄露和不忠实编辑问题。
The MIVE Dataset is a multiple-instance video editing dataset developed by the Korea Advanced Institute of Science and Technology (KAIST) to support the evaluation of multiple-instance video editing tasks. This dataset contains 200 videos covering diverse scenarios and instances, with each video paired with instance-level annotations and masks. The dataset's development process integrates multiple-instance sampling and instance-centric probability redistribution techniques to ensure the accuracy and faithfulness of video editing operations. Primarily utilized in the field of multiple-instance video editing, this dataset is designed to address the issues of editing leakage and unfaithful editing faced by existing methods during multi-object video editing tasks.

- 1MIVE: New Design and Benchmark for Multi-Instance Video Editing韩国科学技术院 · 2024年



