SceneHGN: Hierarchical Graph Networks for 3D Indoor Scene Generation With Fine-Grained Geometry.

Area of Science:

• Computer Graphics
• Artificial Intelligence
• Computational Geometry

Background:

• 3D indoor scenes are crucial for applications like gaming and VR.
• Manual creation of high-quality 3D scenes is time-consuming and requires expertise.
• Existing methods address scene generation partially, focusing on room layout or object geometry separately.

Purpose of the Study:

• To propose a novel method for direct, end-to-end generation of plausible 3D indoor scenes.
• To integrate room layout, object details, and their spatial relationships within a unified framework.
• To overcome limitations of existing partial scene generation approaches.

Main Methods:

• SceneHGN, a hierarchical graph network, models 3D indoor scenes from room to object part levels.
• Functional regions are introduced as intermediate representations to manage complexity.
• A conditional recursive neural network (RvNN)-based variational autoencoder (VAE) generates scene content.

Main Results:

• SceneHGN directly generates complete 3D indoor scenes with fine-grained object geometry and layout.
• The hierarchical graph representation effectively captures inter-level and intra-level relationships.
• Experimental results show superior generation quality compared to existing partial methods.

Conclusions:

• SceneHGN offers a unified approach for 3D indoor scene synthesis, addressing layout and geometry simultaneously.
• The method demonstrates effectiveness in generating detailed and plausible 3D indoor environments.
• SceneHGN is versatile, supporting applications like scene editing, interpolation, and generation from arbitrary boundaries.