Source-linked AI summary
3D-FUTURE: 3D Furniture shape with TextURE
Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, Dacheng Tao
TL;DR
Existing 3D benchmarks often lack detailed geometry, informative textures, and precise image-shape correspondence for fine-grained furniture research. 3D-FUTURE provides a large, richly annotated repository of industrial furniture scenes and evaluates it across several 3D vision tasks, while results show texture recovery remains challenging.
Problem
Existing repository-based 3D benchmarks often provide insufficient geometric detail and uninformative textures, while precise pixel-level 2D-3D alignment is expensive.
Method
3D-FUTURE constructs household interior scenes from industrial furniture shapes with high-resolution textures, rich annotations, and designer-supported furnishing and rendering processes.
Results
3D-FUTURE contains 20,240 realistic synthetic images and 9,992 high-quality 3D CAD furniture shapes, with baselines spanning segmentation, pose, retrieval, reconstruction, and texture recovery.
Takeaways & Limitations
The benchmark is intended to support research on high-quality 3D shape understanding and generation and connect academic study with 3D industrial applications.
Takeaways & Limitations
Visually appealing texture recovery remains challenging: BicycleGAN++ largely ignores semantic parts, while Texture Fields can produce dreamlike textures.
Abstract
from arXiv · showhide
The 3D CAD shapes in current 3D benchmarks are mostly collected from online model repositories. Thus, they typically have insufficient geometric details and less informative textures, making them less attractive for comprehensive and subtle research in areas such as high-quality 3D mesh and texture recovery. This paper presents 3D Furniture shape with TextURE (3D-FUTURE): a richly-annotated and large-scale repository of 3D furniture shapes in the household scenario. At the time of this technical report, 3D-FUTURE contains 20,240 clean and realistic synthetic images of 5,000 different rooms. There are 9,992 unique detailed 3D instances of furniture with high-resolution textures. Experienced designers developed the room scenes, and the 3D CAD shapes in the scene are used for industrial production. Given the well-organized 3D-FUTURE, we provide baseline experiments on several widely studied tasks, such as joint 2D instance segmentation and 3D object pose estimation, image-based 3D shape retrieval, 3D object reconstruction from a single image, and texture recovery for 3D shapes, to facilitate related future researches on our database.
1 Introduction
Existing 3D benchmarks often lack detailed geometry and informative textures, especially for household furniture. 3D-FUTURE addresses this gap with industrial furniture shapes, rich annotations, designer-developed scenes, and baseline tasks for 3D research.
- Existing 3D CAD datasets mainly use online repositories, producing shapes with limited geometry and uninformative or missing textures.Household furniture in these datasets is also described as outdated and unsuitable for fine-grained industrially related research.
- 3D-FUTURE provides modern industrial furniture shapes with fine-grained geometry and attributes covering category, style, theme, and material.The shapes are used in modern industrial production and include informative textures.
- The dataset includes instance segmentation, 6DoF pose, and camera field-of-view annotations alongside textured 3D furniture.These annotations support image-based 3D understanding and interior-scene analysis.
- Experienced designers developed or reviewed the room scenes, while a furnishing suit composition platform was built to support efficient delicate-room design.The platform recurrently recommends visually matched furniture as part of the data-creation process.
- Baseline experiments cover segmentation and pose estimation, image-based shape retrieval, single-image reconstruction, and texture recovery.The experiments are intended to facilitate future research using the database.
2 Related Work
Related 3D benchmarks include model repositories, image-shape datasets, and RGB-D scene collections, but they differ in texture quality, alignment precision, and realism. 3D-FUTURE emphasizes textured industrial furniture, shared rendering information, and designer-selected viewpoints.
- Benchmark Comparison: Table 1 compares representative benchmarks using dimensions such as instances, alignments, texture informativeness, and synthetic imagery.Its notes distinguish raw or pseudo shape correspondences and identify 3D-FUTURE as household-specific industrial furniture data.
- 3D Models: ShapeNet and related repositories provide large collections of 3D models with semantic annotations organized around taxonomies and object properties.ShapeNetCore and ShapeNetSem contain 51,300 and 12,000 models, respectively.
- 3D Models: Image-shape benchmarks such as PASCAL3D+, ObjectNet3D, and Pix3D associate images with 3D shapes, but alignments may be pseudo or expensive to annotate precisely.Pseudo alignments can overlook local shape details, while pixel-level 2D-3D alignment requires substantial labor.
- RGB-D Scenes: RGB-D datasets contribute real-world indoor scenes with depth, surface-normal, semantic, and bounding-box annotations.Examples include NYU Depth V2 and SUN RGB-D, which provide dense labels or 3D bounding boxes.
- RGB-D Scenes: Unlike InteriorNet and Structured3D, 3D-FUTURE provides textured furniture shapes, shared 6DoF pose and camera FoV, and designer-suggested viewpoints.The viewpoints are selected so rendered images preserve the room's overall design idea.
3 Data Acquisition Process
3D-FUTURE combines industrial furniture CAD assets, designer-created interior scenes, and a furnishing-suit composition pipeline that generates and filters aesthetic designs. Its DFSM and decision-tree FSC use visual embeddings and annotated attributes to support furniture compatibility modeling and dataset construction.
- Dataset construction: The construction pipeline starts from a pool of industrial CAD furnishing and interior-finish models enriched with multiple textures, materials, and attributes.Attributes include theme color, style, material, brand, real-world size, and a 500-category, five-level taxonomy.
- Design generation: The pipeline automatically creates interior designs, whose rendered images are manually screened for visual appeal before inclusion in the dataset.The described selection process retained 15,240 visually appealing designs.
- Deep visual embedding: DFSM learns deep visual embeddings through a shared visual embedding network and two transformer-based tasks: mask prediction and compatibility scoring.The model uses a visual embedding network with two transformer encoders and combines margin ranking with hard-sample mining and a classification-loss variant.
- Furnishing suit composition: FSC trains on ordered furnishing suits by predicting the next item from preceding items and constructing candidate sets with style-, color-, and material-matched negatives.Hard-sample mining later selects negatives from top recommendations to refine the visual embedding network and decision tree.
- Furnishing suit composition: The compatibility task represents each candidate suit with a transformer start-token feature, then applies fully connected layers and a sigmoid to produce a suit-quality score.The learned score estimates the compatibility of the candidate furnishing suit.
- Decision-tree FSC: The decision-tree FSC represents each item with six attributes, including learned visual information, and uses labeled positive and negative suits to learn attribute-based matching patterns.Discrete attributes are one-hot encoded, while theme color and real-world size are discretized with k-means before encoding.
4 Properties of 3D-FUTURE
3D-FUTURE combines realistic interior designs with detailed, textured furniture, verified attributes, and 2D–3D annotations across a large household-scale repository. Its organization supports retrieval, recognition, pose estimation, and other 3D vision studies.
- Photo-realistic Synthetic Images: 20,240 synthetic images represent 20,240 interior designs, including 5,000 experienced designs and 15,240 automatically created aesthetic designs.Professional designers suggest camera positions to obtain the best viewpoint for each room.
- Benchmark Tasks: The repository includes baseline experiments for joint instance segmentation and pose estimation, image-based shape retrieval, reconstruction, and texture recovery.Figures describe confusion matrices, AP and AOS histograms, a joint segmentation–pose network, and cross-domain retrieval using non-parametric softmax loss.
- 2D-3D Alignments: 3D-FUTURE provides 9,992 unique 3D shapes and 20,240 scene images, yielding 37,441 cropped image–shape pairs with slight occlusions.The dataset also includes 6DoF pose annotations and instance semantic labels for 34 categories and ten super-categories.
- High-quality Shapes with Informative Textures: 3D-FUTURE contains detailed furniture with informative textures in varied styles, including intricate carvings and models used for modern industrial production.Compared with ShapeNetCore and ModelNet40, its shapes show uniform distributions of vertices and faces without extremely low-resolution examples.
- Fine-Grained Attributes: For each textured shape, professional designers verify attributes spanning 34 categories, 8 super-categories, 19 styles, 15 materials, and 16 themes.These attributes are demonstrated valuable for interior designs and content understanding by industrial productions.
5 Baseline Experiments
3D-FUTURE is evaluated through baseline experiments spanning recognition, segmentation and pose estimation, retrieval, reconstruction, and texture synthesis. Results expose both useful benchmark difficulty and unresolved challenges in detailed furniture understanding.
- Baseline Experiments: The benchmark evaluates shape recognition, joint instance segmentation and pose estimation, image-based retrieval, reconstruction, and texture synthesis.The data split contains 6,699 training shapes, 3,293 test shapes, 14,761 training images, and 5,479 test images.
- Shape Recognition: 3D-FUTURE supports fine-grained furniture recognition because its shapes require modeling local and global geometric details beyond ShapeNet and ModelNet.MVCNN and PointNet++ are used as recognition baselines.
- Image-based 3D Shape Retrieval: 23.4% Top1@R versus 40.6% Top3@R indicates substantial retrieval ambiguity among furniture with similar shape characteristics.The retrieval experiment trains on 31,444 image-shape pairs and evaluates on 5,994 pairs.
- Instance Segmentation and Pose Estimation: 43% mean AVP is achieved for 3D pose estimation, while the AOS–AP gap indicates that estimated rotations can still improve.The baseline obtains 0.55 mean AP for instance segmentation; pose quality is analyzed across occlusion levels and 6DoF annotations.
- Single-View 3D Object Reconstruction: Pixel2Mesh is more robust for single-view reconstruction, but all evaluated methods struggle with shapes containing many geometric details.Evaluation uses IoU, Chamfer Distance, and F-score on randomly rendered 256 × 256 images.
- Texture Synthesis for 3D Shapes: BicycleGAN++ scores higher on FID and Feat1, whereas Texture Fields performs better on SSIM and L1, revealing different texture-synthesis strengths.Qualitative results show BicycleGAN++ mainly learns color, while Texture Fields partially preserves structure but can produce dreamlike textures.
6 Conclusion
The paper introduces 3D-FUTURE as a large-scale household benchmark combining realistic synthetic scenes, detailed industrial furniture, informative textures, and rich 2D–3D annotations. Its experiments establish baselines intended to support future research in 3D understanding and generation.
- 6 Conclusion: 3D-FUTURE contains 20,240 realistic synthetic images and 9,992 high-quality 3D CAD furniture shapes.The benchmark is specific to household scenarios and includes rich 3D and 2D annotations.
- 6 Conclusion: Its features include designer-created interior designs, photo-realistic renderings, 2D–3D alignments, and industrial furniture with informative textures.These properties are presented as the benchmark’s principal distinguishing features.
- 6 Conclusion: The benchmark is intended to facilitate high-quality 3D shape understanding and generation, create research opportunities in 3D vision, and bridge academic study with industrial applications.The paper presents its experiments as baselines for future research using the database.