BL30K

BL30K 是使用 Blender 和 ShapeNet 数据渲染的合成数据集。我们将数据集分成六段，每段大约有 5K 视频。这些视频的组织格式与 DAVIS 和 YouTubeVOS 类似，因此可以直接使用这些数据集的数据加载器。每个视频长160帧，每帧分辨率为768*512。每个视频有 3-5 个对象，每个对象都有一个随机平滑的轨迹——我们试图以贪婪的方式优化轨迹，以最小化对象交叉（不能保证），但仍然可能出现遮挡（现实中经常发生）。有关详细信息，请参阅 MiVOS。

BL30K is a synthetic dataset rendered using Blender and ShapeNet data. We partition the dataset into six splits, each comprising approximately 5K videos. The videos follow the same organizational format as DAVIS and YouTubeVOS, so their existing data loaders can be directly adopted. Each video consists of 160 frames, with each frame having a resolution of 768 × 512. Each video contains 3 to 5 objects, each with a randomly smoothed trajectory. We sought to optimize the trajectories in a greedy fashion to minimize object intersections (this is not guaranteed), yet occlusion may still arise, which is a common occurrence in real-world scenarios. For further details, please refer to MiVOS.

搜集汇总

背景与挑战

背景概述

BL30K是一个合成的视频数据集，包含约30K视频，用于视频目标分割任务，视频格式与DAVIS和YouTubeVOS兼容，每个视频包含3-5个对象，具有随机平滑的轨迹。