BL30K 视频对象分割数据集

BL 全称 Blender，BL30K 是一个合成视频数据集，视频由 Blender（3D 创作软件）采用 ShapeNet 数据集的数据进行合成。该数据集有 6 个部分，每个部分大约有 5 千个视频，每个视频有 160 帧，每帧的分辨率为 768*512 。每段视频有 3—5 个物体，每个物体都有一个随机的平滑轨迹——研究人员试图优化轨迹，尽量减少物体的交叉（不保证），闭塞仍可能存在（现实中经常发生）。这些视频的组织形式与 DAVIS 和 YouTubeVOS 类似，因此可以直接使用这些数据集的数据加载器。详情参见 MiVOS 。

BL is short for Blender. BL30K is a synthetic video dataset generated by Blender (a 3D creation software) using data from the ShapeNet dataset. This dataset is divided into 6 subsets, each containing approximately 5,000 videos. Each video has 160 frames, with a resolution of 768×512 per frame. Each video contains 3 to 5 objects, each with a random smooth trajectory. Researchers attempted to optimize the trajectories to minimize object overlaps (though full avoidance is not guaranteed), and occlusions may still occur, which is a common phenomenon in real-world scenarios. The organization of these videos is similar to that of DAVIS and YouTubeVOS, so their existing data loaders can be directly applied to this dataset. For more details, please refer to MiVOS.

搜集汇总

背景与挑战

背景概述

BL30K是一个基于Blender软件合成的视频对象分割数据集，包含约3万个视频（每段160帧、分辨率768×512），每个视频包含3-5个具有随机平滑轨迹的物体。其数据组织形式与DAVIS和YouTubeVOS兼容，可直接使用现有数据加载器。