DeepFurniture Dataset

A large-scale dataset for furniture understanding, featuring photo-realistic rendered indoor scenes with high-quality 3D furniture models. The dataset contains about 24k indoor images, 170k furniture instances, and 20k unique furniture identities, all rendered by the leading industry-level rendering engines in Kujiale.

This dataset is introduced in our paper: Furnishing Your Room by What You See: An End-to-End Furniture Set Retrieval Framework with Rich Annotated Benchmark Dataset

Key Features

• Photo-Realistic Rendering: All indoor scenes are rendered using professional rendering engines, providing realistic lighting, shadows, and textures
• High-Quality 3D Models: Each furniture identity is derived from a professional 3D model, ensuring accurate geometry and material representation
• Rich Annotations: Hierarchical annotations at image, instance, and identity levels

Dataset Overview

DeepFurniture provides hierarchical annotations at three levels:

• Image Level: Professional rendered indoor scenes with scene category and depth map
• Instance Level: Bounding boxes and per-pixel masks for furniture instances in scenes
• Identity Level: High-quality rendered previews of 3D furniture models.

Statistics

• Total scenes: ~24,000 photo-realistic rendered images
• Total furniture instances: ~170,000 annotated instances in scenes
• Unique furniture identities: ~20,000 3D models with preview renderings
• Categories: 11 furniture types
• Style tags: 11 different styles

Categories

1. Cabinet/Shelf
2. Table
3. Chair/Stool
4. Lamp
5. Door
6. Bed
7. Sofa
8. Plant
9. Decoration
10. Curtain
11. Home Appliance

Style Tags

1. Modern
2. Country
3. European/American
4. Chinese
5. Japanese
6. Mediterranean
7. Southeast-Asian
8. Nordic
9. Industrial
10. Eclectic
11. Other

Dataset Structure

The dataset is organized in chunks for efficient distribution:

data/
├── scenes/           # Photo-realistic rendered indoor scenes
├── furnitures/       # High-quality 3D model preview renders
├── queries/          # Query instance images from scenes
└── metadata/         # Dataset information and indices
    ├── categories.json       # Furniture category definitions
    ├── styles.json          # Style tag definitions
    ├── dataset_info.json    # Dataset statistics and information
    ├── furnitures.jsonl     # Furniture metadata
    └── *_index.json        # Chunk index files

Using the Dataset

1. Download and Extraction

# Clone the repository
git lfs install  # Make sure Git LFS is installed
git clone https://huggingface.co/datasets/byliu/DeepFurniture

# [optional] Extract the dataset (or load the dataset using the compressed files, TBD) 
./uncompress_dataset.sh -s ./DeepFurniture/data -t ./regular_dataset

2. Data Format

Scene Data

• Image: RGB images in JPG format
• Depth: Depth maps in PNG format
• Annotation: JSON files containing:

  {
    "instances": [
      {
        "numberID": 1,
        "boundingBox": {
          "xMin": int,
          "xMax": int,
          "yMin": int,
          "yMax": int
        },
        "styleIDs": [int],
        "styleNames": [str],
        "segmentation": [int], # COCO format RLE encoding
        "identityID": int,
        "categoryID": int,
        "categoryName": str
      }
    ]
  }
  

Furniture Data

• Preview images of 3D models in JPG format
• Metadata in JSONL format containing category and style information

Query Data

• Cropped furniture instances from scenes
• Filename format: [furnitureID]_[instanceIndex]_[sceneID].jpg

3. Loading the Dataset

from deepfurniture import DeepFurnitureDataset

# Initialize dataset
dataset = DeepFurnitureDataset("path/to/regular_dataset")

# Access a scene
scene = dataset[0]
print(f"Scene ID: {scene['scene_id']}")
print(f"Number of instances: {len(scene['instances'])}")

# Access furniture instances
for instance in scene['instances']:
    print(f"Category: {instance['category_name']}")
    print(f"Style(s): {instance['style_names']}")

Benchmarks

This dataset supports three main benchmarks:

1. Furniture Detection/Segmentation
2. Furniture Instance Retrieval
3. Furniture Retrieval

For benchmark details and baselines, please refer to our paper.

License

Apache-2.0

Acknowledgments

• Dataset created by Kujiale
• Rendered using the leading interior design platform in Kujiale
• Thanks to millions of designers and artists who contributed to the 3D models and designs

If you use this dataset, please cite:

@article{liu2024deepfurniture,
  title={Furnishing Your Room by What You See: An End-to-End Furniture Set Retrieval Framework with Rich Annotated Benchmark Dataset},
  author={Liu, Bingyuan and Zhang, Jiantao and Zhang, Xiaoting and Zhang, Wei and Yu, Chuanhui and Zhou, Yuan},
  journal={arXiv preprint arXiv:1911.09299},
  year={2020}
}