FurniScene: A Large-scale 3D Room Dataset with Intricate Furnishing Scenes

TL;DR

FurniScene introduces a large-scale, meticulously designed 3D indoor scene dataset with extensive furniture diversity and rich annotations, addressing realism and diversity gaps in prior datasets. It then presents TSDSM, a two-stage diffusion framework that first generates a furniture list from text and then constructs layout through diffusion-guided retrieval and placement, achieving superior realism on standard metrics. The work includes thorough data acquisition, annotation, and augmentation pipelines, and demonstrates strong qualitative and quantitative gains over baselines. Overall, FurniScene and TSDSM advance scalable, realistic indoor scene synthesis and understanding, with practical implications for gaming, VR, and interior design research.

Abstract

Indoor scene generation has attracted significant attention recently as it is crucial for applications of gaming, virtual reality, and interior design. Current indoor scene generation methods can produce reasonable room layouts but often lack diversity and realism. This is primarily due to the limited coverage of existing datasets, including only large furniture without tiny furnishings in daily life. To address these challenges, we propose FurniScene, a large-scale 3D room dataset with intricate furnishing scenes from interior design professionals. Specifically, the FurniScene consists of 11,698 rooms and 39,691 unique furniture CAD models with 89 different types, covering things from large beds to small teacups on the coffee table. To better suit fine-grained indoor scene layout generation, we introduce a novel Two-Stage Diffusion Scene Model (TSDSM) and conduct an evaluation benchmark for various indoor scene generation based on FurniScene. Quantitative and qualitative evaluations demonstrate the capability of our method to generate highly realistic indoor scenes. Our dataset and code will be publicly available soon.

Figures (6)

• Figure 1: The pipeline of building FurniScene. Our data collection framework consists of the following steps: purchase complete SketchUp indoor scenes from interior designers; extract CAD model and align coordinate axis in 3DMax; render scenes in UE and add semantic labels manually; perform data augmentation to expand the dataset and increase scene diversity; point cloud generation.
• Figure 2: Statistics on the dataset. The darker color represents the 3D-FRONT and the brighter color represents our FurniScene. a) shows the total count of object categories present in each room type, b) illustrates the average number of objects in each room type, and c) displays the distribution of the top 50 most frequently occurring objects in FurniScene.
• Figure 3: Visualization of ScanNet ScanNet, Scan2CAD Scan2CAD, OpenRooms OpenRooms, 3D-FRONT 3D-FRONT, and FurniScene. ScanNet contains a wealth of ornaments, but it does not separate the mesh. The scene in 3D-FRONT is complete, but there are no small ornaments; FurniScene is diverse, rich in detail, and has a lot of ornaments.
• Figure 4: Model architecture. Firstly, FLGM generates a furniture list based on text prompt. Subsequently, FRS utilizes this furniture list to retrieve corresponding furniture models from our furniture database, which are then passed to LGM for the generation of layout information for these models.
• Figure 5: Examples of rooms in FurniScene. The left column displays the unrendered complete layout information of the room. The middle column showcases multiple rendered images of the room after UE rendering. The right column displays the CAD models contained in the room, each of which boasts intricate geometric details. It is worth noting that FurniScene includes numerous furnishings such as books, bottles, clocks, vases, etc.