HQ-YTVIS

对于HQ-YTVIS，我们按照YTVIS中的比率将原始YTVIS训练集 (2238视频) 分为新的火车 (1678视频，75%)，val (280视频12.5%) 和测试 (280视频12.5%) 集。VMT通过迭代训练对HQ-YTVIS火车子集上的遮罩注释进行自我校正，而较小的val和test集则由人工注释器仔细检查并重新标记，以确保较高的遮罩边界质量。

For HQ-YTVIS, we split the original YTVIS training set (2238 videos) into three new subsets according to the ratio specified in YTVIS: the training split (1678 videos, 75%), validation (val, 280 videos, 12.5%), and test (280 videos, 12.5%). VMT conducted self-correction on the mask annotations of the HQ-YTVIS training subset via iterative training. The smaller val and test sets were carefully inspected and re-annotated by human annotators to ensure high-quality mask boundaries.

搜集汇总

背景与挑战

背景概述

HQ-YTVIS是基于YTVIS训练集重新划分的视频实例分割数据集，通过VMT方法自动校正和人工精细标注，以提高遮罩边界质量。该数据集由苏黎世联邦理工学院等机构于2022年发布，相关论文和首页可公开访问。