kubric

Why we included this project

Video understanding models trained on toy datasets tend to fall apart on real footage, and real-world video rarely comes with the clean pixel-level labels you need for supervision. Kubric sidesteps that by generating semi-realistic synthetic multi-object scenes on demand, and the pipeline computes instance segmentation masks, depth maps, and optical flow as part of the output. It leans on pybullet for physics and Blender for rendering, but the code is modular enough that you could plug in a different renderer, and the whole thing runs inside Docker. The bundled examples walk through setting up a scene, keyframing object motion, and exporting annotated frames in just a few lines. If you need controllable, scalable video data for benchmarks or pretraining without hand-labeling, it does the job.