Infinigen Indoors: Photorealistic Indoor Scenes using Procedural Generation

Alexander Raistrick,Lingjie Mei,Karhan Kayan,David Yan,Yiming Zuo,Beining Han,Hongyu Wen,Meenal Parakh,Stamatis Alexandropoulos,Lahav Lipson,Zeyu Ma,Jia Deng

2024-06-18

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 <a class="link-external link-https" href="https://infinigen.org" rel="external noopener nofollow">this https URL</a> for code and videos.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

This paper introduces a system called Infinigen Indoors, which is a procedural generator for indoor scenes based on Blender, capable of creating realistic indoor environments. Unlike the previous Infinigen system (which mainly focuses on natural scenes), Infinigen Indoors extends to indoor scenes, including a diverse procedural generation library for furniture, architectural elements, appliances, and other everyday objects. Additionally, it introduces a constraint-based layout system, where users can express various constraints on scene layout using a domain-specific language and generate scene layouts that satisfy these constraints to the maximum extent.

The system also provides an export tool that allows the generated 3D objects and scenes to be directly used for training embodied agents in real-time simulators like NVIDIA Omniverse and Unreal Engine. Infinigen Indoors is open-source and entirely based on mathematical rules for procedural generation without the need for external static assets.

The main contributions of this paper include:
1. Procedural generation library for indoor assets.
2. Constraint-based layout system, allowing users to specify constraints such as symmetry, spatial relationships, quantity, physical stability, and accessibility through a Python API.
3. A tool for exporting the generated scenes to real-time simulators.

This constraint-based system enhances the usability and customization of the system, allowing users to conveniently express generation goals without worrying about implementation details. Furthermore, the generated data can be used for domain randomization in simulation environments and for training models in tasks such as shadow removal and occlusion boundary detection, demonstrating the effectiveness of the data and the unique customization capability of the system.

The paper also compares with existing datasets and generators, and discusses related work such as real-world datasets, synthetic indoor datasets, and object arrangement and layout generation methods. What sets Infinigen Indoors apart is its complete procedural nature, high customizability, and integrated constraint-solving capability.