High quality data, especially large-scale labeled data plays a crucial role in training models for a wide range of computer vision tasks. Synthetic data from computer graphics is one of the most promising approach for generating unlimited high-quality labeled vision data, but existing freely available synthetic datasets are restricted to a narrow range of objects and shapes, thereby failing to capture the diversity and complexity of the real-world objects.
To bridge this gap, in a new paper titled Infinite Photorealistic Worlds using Procedural Generation, a Princeton University research team presents Infinigen, a procedural photorealistic 3D scenes generator that is capable to generate unlimited and diverse training data of the natural world, substantially expands the coverage of existing synthetic data.
Infinigen is entirely procedural that can generate infinitely quantity of shapes, textures, materials, and scene compositions from scratch. It achieves high photorealism by procedurally generating both coarse structures and fine details in geometry and texture and all geometric details in Infinigen are real.
In terms of model architecture, Infinigen is built upon Blender, a graphics system for procedural generation. To utilize the useful primitives from Blender, the team designs and implements a library of procedural rules to expand the coverage of natural objects and scenes. They further develop utilities to facilitate creation of procedural rules, and it can automatically converts Blender node graphs to Python code; and they also develop utilities to render synthetic images with ground truth labels, including depth, occlusion boundaries, bounding boxes, optical flow, surface normals, object category, and instance segmentation.
In their empirical study, the team performs extensive experiments to validate the quality of the generated synthetic data and compared Infinigen to existing synthetic datasets or generators. The results demonstrate its capability to generate photorealistic infinite original assets and scenes of the real world without external sources.
Infinigen is free and open-source, the researchers claim they intend to make Infinigen a living project through collaboration with the whole community as well as expanding its coverage to everything in the real world.
Author: Hecate He | Editor: Chain Zhang
We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.