Data, especially large scale labeled data, has been a critical driver of progress in computer vision. At the same time, many tasks remain starved of high-quality data. Synthetic data from computer graphics is a promising solution to this data challenge, but sees limited use, which we hypothesize is due to limited diversity of assets.
We propose a synthetic data generator which can produce infinite photorealistic 3D scenes of the natural world. Our system is entirely procedural: every asset, from shape to texture, is generated from scratch via randomized mathematical rules. This results in a limitless variety of 3D assets, with rich compositionality and pixel-level 3D detail. Moreover, it means that all assets are original, relying on no external source, and can be released for free as open source code. We provide such procedural rules for a wide range of natural objects and scenes, including plants, animals, terrain and lighting.
Procedural modeling requires significant programming effort, and generating full scenes poses a large computational challenge. To this end, we provide tools and frameworks to accelerate procedural rule creation, and enable large-scale rendering of complex procedurally generated scenes.
In all, anyone can freely use our system to generate unlimited high-quality training data for a wide range of computer vision tasks, including object detection, semantic segmentation, optical flow and 3D reconstruction. We expect this to be a useful tool for computer vision research and beyond.