3D Indoor Scene Synthesis for E‑Commerce: Spatial Perception and Layout Generation

This article introduces the application of 3D scene synthesis technology in the e‑commerce domain, especially for home‑decoration. It explains how realistic indoor scenes enable interactive product viewing, improve shopping experience, and reduce return rates.

Two main categories of indoor‑scene generation algorithms are described: traditional rule‑based/optimization methods and deep‑learning approaches such as diffusion models and large language models (e.g., MVDiffusion, LayoutGPT). The paper emphasizes the need to satisfy physical, functional and aesthetic constraints while ensuring that the objects correspond to actual product models.

To handle massive product catalogs, the authors propose a parameterized representation based on 3D bounding boxes. Each object is defined by a fixed point (bottom‑center or top‑center), orientation, and unknown dimensions (length, width, height). Linear constraints encode collisions, floating avoidance, wall alignment, passage clearance and access space.

Collision detection is performed with the Separating Axis Theorem (SAT) on oriented bounding boxes. Floating problems are solved by fixing the top‑center for ceiling objects. Wall‑adjacent furniture uses a “back_center” fixed point to keep the contact edge aligned.

Functional constraints such as sufficient passage area and furniture access space are modeled by enlarging bounding boxes of existing objects or inserting virtual obstacles. Aesthetic constraints are addressed by maintaining orderliness and consistent style.

The workflow starts from a layout design, then uses spatial‑perception rules to retrieve suitable product models, place them according to the parameterized equations, and finally refine the scene. The authors discuss future directions, including end‑to‑end generation from scratch and tighter integration with LLM‑driven design agents.