LVOS
收藏资源简介:
LVOS数据集是复旦大学创建的,专注于长期视频对象分割的基准数据集。该数据集包含220个视频,总时长421分钟,平均每个视频时长1.59分钟,远超现有短期视频数据集。LVOS数据集通过半自动标注流程,实现了所有帧的高质量密集标注,包含27个类别,其中7个为未见类别,以评估模型的泛化能力。该数据集旨在解决长期视频中对象频繁出现和消失的问题,以及跨时间相似对象的混淆问题,为视频理解和编辑、增强现实等应用提供支持。
The LVOS dataset is a benchmark dedicated to long-term video object segmentation, created by Fudan University. It contains 220 videos with a total duration of 421 minutes, averaging 1.59 minutes per video, which far exceeds the scale of existing short-term video datasets. All frames in the dataset are annotated with high-quality dense labels via a semi-automatic annotation pipeline, covering 27 categories in total, among which 7 are unseen categories for evaluating the generalization ability of models. This dataset aims to tackle the issues of frequent appearance and disappearance of objects in long-term videos, as well as the confusion of similar objects across time, providing support for applications such as video understanding and editing, augmented reality, and other relevant fields.




