Infinigen Indoors: Photorealistic Indoor Scenes using Procedural Generation

TL;DR

Infinigen Indoors addresses the need for high-fidelity synthetic indoor data for training embodied AI by extending the fully procedural Infinigen framework to indoor environments. It presents a Blender-based pipeline with 79 asset generators and 30 material generators, a constraint-based Domain-Specific API, and a simulated-annealing arrangement solver that optimizes floor plans and object layouts using a mix of discrete and continuous moves, with acceptance governed by $p(s'|s) = \min \left [ \exp \left ( \frac{l(s) - l(s')}{\tau} \right ), 1 \right ]$ and a temperature schedule. Key contributions include a 1058-node constraint program example with 11 hard and 25 soft terms, a floorplan generator based on Mondrian Process and PCFGs, and a one-click USD export path to Omniverse and Unreal for real-time simulation. Experimental results show solver performance gains, effective shadow removal data generation, and superior occlusion boundary estimation when trained on Indoors data, with the work released under BSD for broad open-source use.

Abstract

We introduce Infinigen Indoors, a Blender-based procedural generator of photorealistic indoor scenes. It builds upon the existing Infinigen system, which focuses on natural scenes, but expands its coverage to indoor scenes by introducing a diverse library of procedural indoor assets, including furniture, architecture elements, appliances, and other day-to-day objects. It also introduces a constraint-based arrangement system, which consists of a domain-specific language for expressing diverse constraints on scene composition, and a solver that generates scene compositions that maximally satisfy the constraints. We provide an export tool that allows the generated 3D objects and scenes to be directly used for training embodied agents in real-time simulators such as Omniverse and Unreal. Infinigen Indoors is open-sourced under the BSD license. Please visit https://infinigen.org for code and videos.

Figures (25)

• Figure 1: Random, non cherry-picked sample of images generated by our system. From top left to bottom right, we show images from dining rooms, bathrooms, living rooms and kitchens. Please see Appendix \ref{['sec:rand_sample']} for an extended random sample.
• Figure 2: Each image (a) is rendered from a mesh (b), from which we can also extract Depth (c), Surface Normals (d), Occlusion Boundaries (e), Segmentation (f), Bounding Boxes (e) and Optical Flow (h), with Albedo (i) from rendering metadata.
• Figure 3: Random samples of procedurally generated doors (top), staircases (middle/bottom) and windows/warehouse shelving (bottom-right).
• Figure 4: Random samples of procedurally generated ovens, dishwasher and sinks (top/middle), living-room furniture (middle) and bathroom fixtures (bottom).
• Figure 5: Random samples of procedurally generated furniture, including sofa, chairs, and beds (top), tables (middle/bottom), and shelves (bottom).