[Paper Review] Automatic Generation of Constrained Furniture Layouts

Paul Henderson, Kartic Subr+1 more|arXiv (Cornell University)|Nov 29, 2017

Advanced Manufacturing and Logistics Optimization43 references14 citations

TL;DR

This paper proposes a data-driven, probabilistic generative model for creating realistic 3D furniture layouts under diverse constraints such as room shape, size, and object placement. By learning joint statistics of object categories and their spatial relationships from a large training dataset, the method enables efficient ancestral sampling (0.04s per layout) and supports flexible constraint handling via rejection sampling, producing layouts indistinguishable from human-designed ones in user studies.

ABSTRACT

Efficient authoring of vast virtual environments hinges on algorithms that are able to automatically generate content while also being controllable. We propose a method to automatically generate furniture layouts for indoor environments. Our method is simple, efficient, human-interpretable and amenable to a wide variety of constraints. We model the composition of rooms into classes of objects and learn joint (co-occurrence) statistics from a database of training layouts. We generate new layouts by performing a sequence of conditional sampling steps, exploiting the statistics learned from the database. The generated layouts are specified as 3D object models, along with their positions and orientations. We show they are of equivalent perceived quality to the training layouts, and compare favorably to a state-of-the-art method. We incorporate constraints using a general mechanism -- rejection sampling -- which provides great flexibility at the cost of extra computation. We demonstrate the versatility of our method by applying a wide variety of constraints relevant to real-world applications.

Motivation & Objective

• To address the challenge of automatically generating realistic, controllable 3D indoor furniture layouts at scale for virtual environments.
• To support a wide range of user-specified constraints, including room geometry, object placement, and spatial relationships.
• To ensure generated layouts are perceptually realistic and comparable to human-designed layouts.
• To enable efficient, interpretable, and controllable layout generation using learned probabilistic statistics.
• To provide a flexible framework that supports both a priori constraint specification and parameter editing for customization.

Proposed method

• The method models room composition as a directed acyclic graph of object categories, learning conditional probability distributions over object types, positions, and orientations from a database of 250,000 human-designed layouts.
• It performs ancestral sampling using learned multivariate normal distributions for continuous variables (position, orientation) and categorical distributions for object classes.
• Constraints are enforced via rejection sampling, allowing flexible support for multiple simultaneous constraints such as room shape, door/window placement, and object-specific rules.
• The model incorporates high-order spatial relationships (e.g., motifs like chairs around a table, abutments like cabinets in a row) through learned co-occurrence statistics.
• Interpretability is enhanced by visualizing learned parameters, such as heatmaps of preferred placement locations and orientation biases for each object class.
• Efficient parameter editing is supported by modifying learned distributions (e.g., adjusting padding or edge probability) without retraining.

Experimental results

Research questions

• RQ1 Can a data-driven, probabilistic model generate realistic 3D furniture layouts that match human-designed quality?
• RQ2 To what extent can the model support diverse, user-specified constraints such as room shape, object placement, and spatial relationships?
• RQ3 How efficiently can the model generate valid layouts under constraints, and can it produce layouts not present in the training data?
• RQ4 How do users perceive the realism of generated layouts compared to state-of-the-art methods?
• RQ5 Can learned model parameters be interpreted and edited meaningfully to guide layout generation?

Key findings

• The model generates valid layouts in cases where no training example satisfies the given constraints, demonstrating generalization beyond the training distribution.
• Layout generation is highly efficient, with a median sampling time of 0.04 seconds per room, enabling real-time application.
• User studies show that generated layouts are perceived as equally realistic as human-designed layouts and are preferred over the state-of-the-art method [WSCR18].
• The model learns human-interpretable spatial patterns: for example, showers are placed in corners (99% probability), and dining tables are centrally located.
• Objects like office chairs and tripods are less likely to be placed at room edges (0.74 and 0.49 probability, respectively), aligning with real-world usage.
• The model supports efficient parameter editing—e.g., modifying edge probability or orientation bias—enabling direct manipulation of layout characteristics without rejection sampling.