Source-linked AI summary
3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics
Huan Fu, Bowen Cai, Lin Gao, Lingxiao Zhang, Jiaming Wang Cao Li, Zengqi Xun, Chengyue Sun, Rongfei Jia, Binqiang Zhao, Hao Zhang
TL;DR
Existing indoor-scene datasets are limited by reconstruction quality, incomplete synthetic scene packages, or unavailable assets. 3D-FRONT constructs a large repository of professionally designed, style-compatible furnished scenes using compatibility recommendation and layout optimization. It provides 18,968 furnished rooms, high-quality textured furniture, rendering support, and applications in texture synthesis and furniture arrangement.
Problem
Existing indoor-scene datasets can have low-quality reconstructed meshes or lack complete public scene packages containing layouts, design ideas, and involved CAD objects.
Method
3D-FRONT builds furnished synthetic scenes from professional house designs using compatibility recommendation, layout optimization, verification, and practical viewpoint assignment.
Results
18,968 furnished rooms and professionally designed layouts make 3D-FRONT the largest publicly available collection, with style-compatible furniture, high-quality textures, and rendering support.
Takeaways & Limitations
The dataset supports applications in texture synthesis and coherent furniture arrangement that are not well supported by other publicly available datasets.
Takeaways & Limitations
Acquired-scene datasets remain constrained by challenging depth-camera reconstruction, which can reduce mesh quality and annotation precision.
Abstract
from arXiv · showhide
We introduce 3D-FRONT (3D Furnished Rooms with layOuts and semaNTics), a new, large-scale, and comprehensive repository of synthetic indoor scenes highlighted by professionally designed layouts and a large number of rooms populated by high-quality textured 3D models with style compatibility. From layout semantics down to texture details of individual objects, our dataset is freely available to the academic community and beyond. Currently, 3D-FRONT contains 18,968 rooms diversely furnished by 3D objects, far surpassing all publicly available scene datasets. In addition, the 13,151 furniture objects all come with high-quality textures. While the floorplans and layout designs are directly sourced from professional creations, the interior designs in terms of furniture styles, color, and textures have been carefully curated based on a recommender system we develop to attain consistent styles as expert designs. Furthermore, we release Trescope, a light-weight rendering tool, to support benchmark rendering of 2D images and annotations from 3D-FRONT. We demonstrate two applications, interior scene synthesis and texture synthesis, that are especially tailored to the strengths of our new dataset. The project page is at: https://tianchi.aliyun.com/specials/promotion/alibaba-3d-scene-dataset.
1. Introduction
3D-FRONT addresses limitations in existing indoor-scene datasets by providing professionally designed layouts, many furnished rooms, high-quality textured objects, and style-compatible interiors. It also supports rendering and demonstrates scene- and texture-synthesis applications.
- Existing datasets include scanned scenes with noisy, incomplete reconstructions and often low geometric and texture quality.
- 3D-FRONT is a large-scale synthetic indoor-scene repository with professionally designed layouts, object semantics, and 18,968 furnished rooms.
- Its 3D furniture objects have high-quality textures, while recommender-based selection uses expert designs and furniture style compatibility.
- 3D-FRONT offers more professionally designed rooms than its 18,968 furnished-room release, with the total approaching 45,000.
- Trescope enables users to capture desired 2D renderings and annotations, supporting image-driven learning tasks.
- The paper demonstrates texture synthesis and coherent furniture arrangement for empty rooms, applications especially suited to style-consistent, high-quality textured meshes.
2. Related Work
Indoor-scene datasets are divided into acquired scans and designed synthetic scenes, each offering different annotations, scale, and asset availability. 3D-FRONT provides completed scene packages with layouts, design ideas, and involved objects.
- Current 3D-scene datasets mainly arise from scanning and reconstruction or human creation, forming acquired and designed categories.
- Acquired Scenes: Acquired datasets collect RGB-D sequences, reconstruct scenes, and add labels such as pixel annotations, polygons, bounding boxes, layouts, and categories.
- Acquired Scenes: Depth-camera reconstruction remains challenging, so acquired datasets typically have lower mesh quality and may contain imprecise camera-pose or 2D–3D-alignment annotations.
- Designed Scenes: Designed datasets use professional design software, but many large-scale synthetic datasets do not publicly release complete scene packages and their detailed CAD assets.
- Designed Scenes: 3D-FRONT shares the components used to construct houses, including real layouts, interior design ideas, and involved objects.
3. Building 3D-FRONT
3D-FRONT is built by transforming professionally designed houses into furnished, verified scenes through room-suite creation, compatibility recommendation, layout optimization, and viewpoint assignment.
- Building 3D-FRONT: The pipeline starts from about 60K professionally designed houses and 1M CAD meshes, creates room suites, optimizes layouts, verifies designs, and assigns camera viewpoints.
- Room Suite Creation: Room-suite creation represents design ideas through object categories, positions, orientations, sizes, and styles, then recurrently fills rooms with visually matched furniture.
- Room Suite Creation: FSC models visual compatibility using masked-furniture prediction and suite-compatibility scoring based on expert scene designs.
- Room Suite Creation: The improved recommendation stage uses a graph auto-encoder with object embeddings and pairwise visual-appeal edges to rank furniture candidates.
- Layout Optimization and Verification: Layout optimization addresses artifacts such as overlapping furniture by modifying positions under distance, accessibility, wall, focal-point, and collision constraints.
- Layout Optimization and Verification: The energy function runs for up to 50 iterations and takes 10s per room on average.
- Layout Optimization and Verification: Trescope supports interactive review and offline rendering of images, depth, normals, and segmentation for 3D-FRONT scenes.
- Viewpoint Generation: Viewpoint generation transfers knowledge from roughly 5,000 expert-designed rooms to assign practical cameras to each scene.
4. Validation and Assessment
3D-FRONT is assessed through recommender-system metrics, user studies against SUNCG, dataset statistics, and examples demonstrating its furnished layouts and object quality.
- Recommender-system evaluation: 88% of AI-created designs were rated high-quality or customer-preferred, compared with 71% for ordinal-level designers.The comparison reuses professional design ideas, so the authors note it may not be fair to ordinal-level designers.
- Recommender-system evaluation: FSC+GAE generally improves all reported recommender-system metrics over the original FSC.The evaluation uses recommendations from an extremely large pool of about 1M models, after filtering invalid retrieval items by fine-grained category labels.
- User study: 60%–70% of AMT users preferred 3D-FRONT over SUNCG across every assessed quality criterion.The study compared 90 scene pairs and 30 model pairs, with each pair labeled by 20 master-level annotators.
- Dataset properties and applications: 3D-FRONT publicly shares layout semantics, stylistic and texture details, and professionally sourced layout ideas for high-quality indoor-scene modeling.The dataset supports tasks including floorplan synthesis, interior scene synthesis, compatibility prediction, SLAM, reconstruction, and segmentation.
- Dataset properties and applications: Trescope provides lightweight rendering of 2D images and annotations from furnished 3D-FRONT scenes.The tool is intended to simplify rendering for 2D vision studies, while the released repository supports varied scene-analysis and modeling applications.
- Dataset statistics and examples: 6,813 distinct houses contain 44,427 rooms, averaging 6.5 rooms per house.House examples show top-down layouts, contained room types, and the textured objects and high-quality CAD models used in their interiors.
5. Applications
3D-FRONT supports interior scene synthesis and context-aware texture synthesis through professionally designed layouts and style-compatible, textured furniture. Demonstrations show improved plausibility, diversity, and texture richness over alternative training data.
- 3D-FRONT is used to demonstrate interior scene synthesis and object texturing in scene contexts.
- Interior Scene Synthesis: 6,230 bedrooms and 645 living rooms were selected for scene synthesis, with separate training and evaluation splits.The experiment retained rooms no larger than 6 meters in width or length and used 6,070/485 rooms for training and 160/160 for evaluation.
- Interior Scene Synthesis: 64.8% of AMT users preferred scenes synthesized from 3D-FRONT over 35.2% for SUNCG.Users judged scenes generated from randomly chosen empty rooms for plausibility.
- Interior Scene Synthesis: 3D-FRONT-trained synthesis produced richer object variety and more plausible layouts than SUNCG-trained synthesis in qualitative comparisons.The comparison used bedroom and living-room scenes generated from randomly selected empty rooms.
- Texturing 3D Models in Indoor Scenes: Scene texturing guides other objects’ texture generation with VGG features extracted from a randomly chosen object.This extends TM-Net to enforce texture coherence among objects in an indoor scene.
- Texturing 3D Models in Indoor Scenes: TM-Net trained on 3D-FRONT achieved LPIPS 0.289 versus 0.215 for ShapeNet when generating chair textures conditioned on table textures.LPIPS measures diversity of generated textures; an AMT study selected the 3D-FRONT results as richer 61.1% versus 38.9%.
6. Conclusion and future work
3D-FRONT provides a large public collection of professionally designed room layouts instantiated with high-quality textured CAD meshes. Its professional designs, model quality, and style compatibility support applications that other datasets do not adequately support.
- 3D-FRONT offers the largest publicly available collection of professionally designed room layouts instanced with high-quality textured CAD meshes.
- 3D-FRONT surpasses SUNCG in professional layout designs, CAD model quality, and style compatibility.
- These distinctive dataset features enable data-driven applications that other datasets do not well support.The paper identifies interior scene synthesis and object texturing as demonstrated examples.
7. Metrics in “Validation and Assessment”
The paper provides an example to clarify metrics introduced in the validation and assessment section.
- Figure 9 gives an example intended to clarify metrics introduced in Section 4.
8. Other Statistics
The supplementary statistics describe object counts, labels, physical dimensions, and room-conditioned category distributions in 3D-FRONT. These distributions expose relationships between furniture categories and room types.
- Figure 10 reports the distribution of the number of functional furniture objects per room.
- Figure 11 reports annotated object-label distributions corresponding to 3D-FUTURE CAD model categories.
- Figure 13 shows room-conditioned object-category frequencies, with square area denoting frequency and normalization applied per category.
- Children cabinets, bunk beds, and kids beds are more likely in kid rooms, while bookcases and desks are more likely in study rooms.The distributions are presented as a source of design knowledge.
- Figure 12 presents physical sizes across rooms and houses using real-world dimensions in meters.
9. User Studies
The user studies evaluate 3D-FRONT’s scene and model quality against existing datasets, and assess layout and texture synthesis applications. Across the dataset comparison, 3D-FRONT is preferred on the evaluated quality criteria.
- Dataset Quality: 90 scene pairs compared 3D-FRONT and SUNCG on layout plausibility, design quality, texture quality, and style compatibility.The study covered LivingRoom, DiningRoom, and Bed Room scenes, with 20 master-level annotators labeling each pair.
- Model Quality: 30 furniture-model pairs compared SUNCG and 3D-FRONT on texture quality and visual quality.The final scores were calculated from 600 feedback responses.
- Dataset Statistics: The figures also show distributions of annotated instances, object categories by room type, and physical room and house sizes.These visualizations characterize the dataset used in the studies.
- Layout Synthesis: 60 rooms were used to compare layout synthesis models trained on 3D-FRONT and SUNCG.The evaluation studied layout plausibility using 1,200 feedback responses.
- Texture Synthesis: Five DiningRoom corners were used to compare texture synthesis models trained with 3D-FRONT and ShapeNet.Chairs and tables were textured three times per corner with random noises, and texture diversity was evaluated from 100 feedback responses.
- Dataset Quality: 60%–70% of Turkers preferred 3D-FRONT for each assessed dataset quality criterion.The comparisons evaluated layout plausibility, design quality, texture quality, and style compatibility against SUNCG and ShapeNet.
10. More House & Room Examples
The paper provides additional visual examples of houses and rooms to demonstrate the quality of 3D-FRONT. These examples are presented across separate house and room figure groups.
- House Examples: Figures 15–17 present additional house examples from 3D-FRONT.The paper directs readers to zoom in for better viewing.
- Room Examples: Figures 18–19 present additional room examples from 3D-FRONT.The paper directs readers to zoom in for better viewing.