YouTube-VIS 2021

该数据集名为YouTube-VIS 2021，相较于2019版本，它在训练视频数量上有所增加，具体包括2985个训练视频、421个验证视频以及453个测试视频，同时保持了40个对象类别。此外，该数据集的标注质量相较于YouTube-VIS 2019有了显著提升。在规模上，它涵盖了40个对象类别，以及上述提到的视频数量。该数据集的任务是对视频实例进行分割。

This dataset is named YouTube-VIS 2021. Compared to the YouTube-VIS 2019 iteration, it features an increased number of training videos, specifically 2985 training videos, 421 validation videos, and 453 test videos, while retaining 40 object categories. Additionally, the annotation quality of this dataset has been significantly improved relative to YouTube-VIS 2019. In terms of scale, it covers 40 object categories alongside the aforementioned number of videos. The primary task of this dataset is video instance segmentation.

搜集汇总

背景与挑战

背景概述

YouTube-VIS 2021是YouTube-VOS数据集的子集，专注于视频实例分割任务，用于同时检测、分类和跟踪视频中的物体实例。