BL30K Video Object Segmentation Dataset

BL stands for Blender. BL30K is a synthetic video dataset. The videos are synthesized by Blender (3D creation software) using data from the ShapeNet dataset. The dataset has 6 parts, each with about 5,000 videos, each with 160 frames, and the resolution of each frame is 768*512. Each video has 3-5 objects, and each object has a random smooth trajectory - researchers try to optimize the trajectory to minimize the intersection of objects (not guaranteed), and occlusion may still exist (often happens in reality). The organization of these videos is similar to DAVIS and YouTubeVOS, so the data loaders of these datasets can be used directly. For details, see MiVOS.