UVO (Unidentified Video Objects: A Benchmark for Dense, Open-World Segmentation)

UVO 是视频中与开放世界类别无关的对象分割的新基准。除了将问题焦点转移到开放世界设置之外，UVO 的规模要大得多，与 DAVIS 相比，提供的视频大约多 8 倍，与 YouTube-VOS 和 YouTube-VIS 相比，每个视频的掩码（实例）注释多 7 倍。 UVO 也更具挑战性，因为它包含许多具有拥挤场景和复杂背景运动的视频。数据集的一些亮点包括：
高质量实例掩码在 1024 个 YouTube 视频上以 30 fps 密集注释，在 Kinetics 数据集的 10337 个视频上以 1fps 密集注释
开放世界：注释每个视频中的所有对象，平均每个视频 13.5 个对象
对象类别多样：57% 的对象不属于 COCO 类别

UVO is a novel benchmark for open-world category-agnostic object segmentation in videos. Beyond shifting the problem focus to the open-world setting, UVO boasts a significantly larger scale: it offers approximately 8 times more videos than DAVIS, and 7 times more instance mask annotations per video than both YouTube-VOS and YouTube-VIS. Furthermore, UVO is more challenging, as it encompasses numerous videos featuring crowded scenes and complex background movements. Key highlights of the dataset are as follows:
1. High-quality instance masks densely annotated at 30 FPS across 1,024 YouTube videos, and at 1 FPS on 10,337 videos from the Kinetics dataset
2. Open-world annotation: All objects in each video are annotated, with an average of 13.5 objects per video
3. Diverse object categories: 57% of the annotated objects do not fall within COCO categories