Reproduce “Kubric: A scalable dataset generator” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam and 26 more
Kubric provides reusable infrastructure for generating labeled synthetic vision data. Researchers define scenes in Python; Kubric coordinates physics simulation, photorealistic rendering, and data export to produce segmentation, depth, optical-flow, surface-normal, and object-coordinate labels. On Sintel Final, RAFT pretraining with Kubric data yielded 3.02 AEPE, versus 3.76 for FlyingChairs; AutoFlow scored 2.75 with target-tuned parameters.
Key insights
- High-fidelity raytracing through Blender's Cycles engine is critical for generating photo-realistic synthetic data that effectively reduces the sim-to-real gap.
- A unified, open-source framework with standardized outputs (SunDs/TFDS) promotes reproducibility, reusability, and collaborative data generation at scale.
- Fine-grained control over scene parameters enables researchers to systematically test hypotheses, isolate variables, and create targeted benchmarks for specific machine learning challenges.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.