Infinigen

Procedural 3D Scene Generator for Computer Vision Research

Free

Description

Infinigen, developed by Princeton Vision & Learning Lab, is a procedural generator specifically designed for creating 3D scenes optimized for computer vision research. It focuses on generating diverse and high-quality 3D training data automatically. Built upon Blender, Infinigen is offered as a free and open-source tool under the BSD 3-Clause License, encouraging community contributions and ongoing development to expand its capabilities.

The core strength of Infinigen lies in its procedural approach, utilizing randomized mathematical rules to create everything from macro structures to micro details, including shapes and materials, ensuring unlimited variations. It excels in generating natural world elements like plants, animals, terrains, and phenomena such as fire and clouds. A key aspect is its commitment to 'real geometry', avoiding techniques that fake geometric detail, which ensures accurate 3D ground truth. Furthermore, Infinigen automatically produces high-quality annotations for tasks like optical flow, depth estimation, and panoptic segmentation, which are easily customizable.

Key Features

Procedural Generation: Creates assets and scenes using randomized mathematical rules, allowing unlimited variations and user control.
Diverse Asset Library: Generates plants, animals, terrains, and natural phenomena like fire, cloud, rain, and snow.
Real Geometry: Produces detailed, accurate 3D geometry without faking details, ensuring precise ground truth.
Automatic Annotations: Generates high-quality, customizable annotations for various computer vision tasks (optical flow, depth, segmentation, etc.).
Open Source & Free: Available under the BSD 3-Clause License and based on Blender.