SUN-RGBD / README.md
dronefreak's picture
Add dataset card and animated banner (#2)
bea939b
|
Raw History Blame Contribute Delete
9.39 kB
metadata
language:
  - en
license: other
size_categories:
  - 10K<n<100K
task_categories:
  - image-segmentation
  - depth-estimation
  - object-detection
pretty_name: SUN RGB-D
license_name: sun-rgbd-no-license-specified
license_link: https://rgbd.cs.princeton.edu/
tags:
  - rgb-d
  - indoor
  - scene-understanding
  - semantic-segmentation
  - 3d-object-detection
  - scene-classification
  - computer-vision
dataset_info:
  features:
    - name: image
      dtype: image
    - name: depth
      dtype: image
    - name: depth_filled
      dtype: image
    - name: semantic37
      dtype: image
    - name: semantic13
      dtype: image
    - name: scene
      dtype: string
    - name: sensor
      dtype: string
    - name: source
      dtype: string
    - name: path
      dtype: string
    - name: intrinsics
      list: float32
    - name: extrinsics
      list: float32
    - name: objects_2d
      list:
        - name: name
          dtype: string
        - name: bbox
          list: float32
        - name: polygon_x
          list: float32
        - name: polygon_y
          list: float32
    - name: objects_3d
      list:
        - name: name
          dtype: string
        - name: corners_x
          list: float32
        - name: corners_z
          list: float32
        - name: y_min
          dtype: float32
        - name: y_max
          dtype: float32
    - name: annotation_2d_json
      dtype: string
    - name: annotation_3d_json
      dtype: string
  splits:
    - name: train
      num_bytes: 1855639927
      num_examples: 5285
    - name: test
      num_bytes: 1822135402
      num_examples: 5050
  download_size: 3623928174
  dataset_size: 3677775329
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

SUN RGB-D: Indoor RGB-D Scene Understanding Benchmark

Task Dataset Format Classes Splits License

Unofficial redistribution of SUN RGB-D in a viewer-ready Parquet layout: RGB, raw and filled depth, 37- and 13-class semantic labels, scene category, camera parameters, and 2D/3D object annotations in one row per frame.

Dataset Description

Disclaimer

This repository is not an official release of SUN RGB-D.

SUN RGB-D was created by Shuran Song, Samuel P. Lichtenberg and Jianxiong Xiao (Princeton University). It includes images from NYU Depth v2 (Silberman et al.), Berkeley B3DO (Janoch et al.) and SUN3D (Xiao et al.). The original authors retain all copyright and intellectual property rights; this repository does not claim ownership of any images, annotations, or metadata.

No license was specified by the original authors. The project page only asks users to cite the SUN RGB-D paper and the three source datasets. This copy is an unofficial redistribution for research purposes, with full attribution. If you are an author or rights holder and want this repository removed or changed, please open a discussion on this repository and it will be addressed promptly.

The 37- and 13-class semantic labels are the extraction by Ankur Handa (sunrgbd-meta-data, no license specified), made from the official SUNRGBD2Dseg.mat.


Dataset Overview

SUN RGB-D contains 10,335 indoor RGB-D images captured with four different sensors (Intel RealSense, Asus Xtion, Microsoft Kinect v1 and v2), densely annotated with 146,617 2D polygons and 58,657 3D bounding boxes, plus scene categories and room layouts. It is a standard benchmark for indoor semantic segmentation, scene classification, and 3D object detection.


Changes from the Official Release

  • Repacked into Parquet (one row per frame) so the Hugging Face Dataset Viewer and datasets work out of the box, with full browsing, search and filtering.
  • RGB, depth (depth/), filled depth (depth_bfx/), scene category, intrinsics, extrinsics and the final 2D/3D annotations are taken unchanged from the official SUNRGBD.zip.
  • Added the 37- and 13-class semantic label maps from Ankur Handa's extraction. Their order follows the official 5,285 / 5,050 train/test split; alignment was verified during the build (label size equals image size for every frame, and unlabeled pixels match each frame's own seg.mat).
  • objects_2d and objects_3d are parsed from annotation2Dfinal and annotation3Dfinal; the raw JSON is kept in annotation_2d_json and annotation_3d_json.
  • Not included: the fullres/ copies, intermediate annotation versions, and room layouts.
  • No images were added, removed, or modified. No labels were changed. The official split is unchanged.

Dataset Structure

Column Content
image RGB image (resolution depends on the sensor)
depth raw depth, 16-bit PNG (official depth/)
depth_filled hole-filled depth, 16-bit PNG (official depth_bfx/)
semantic37 37-class label map (0 = unlabeled)
semantic13 13-class label map (0 = unknown)
scene scene category (e.g. bedroom, office)
sensor, source kv1 / kv2 / realsense / xtion, and the sub-collection (e.g. NYUdata, b3dodata, sun3ddata)
path original folder inside SUNRGBD.zip
intrinsics 3x3 camera matrix, row-major
extrinsics 3x4 rotation/translation from the latest extrinsics/*.txt
objects_2d list of {name, bbox [x, y, w, h], polygon_x, polygon_y}
objects_3d list of {name, corners_x, corners_z, y_min, y_max} (box footprint + height, as in the toolbox)
annotation_2d_json, annotation_3d_json raw annotation files

Splits: train 5,285 · test 5,050 (10,335 total).

Classes

37-class (semantic37, 1..37): wall, floor, cabinet, bed, chair, sofa, table, door, window, bookshelf, picture, counter, blinds, desk, shelves, curtain, dresser, pillow, mirror, floor_mat, clothes, ceiling, books, fridge, tv, paper, towel, shower_curtain, box, whiteboard, person, night_stand, toilet, sink, lamp, bathtub, bag

13-class (semantic13, 1..13): bed, books, ceiling, chair, floor, furniture, objects, picture, sofa, table, tv, wall, window

Usage

from datasets import load_dataset

ds = load_dataset("dronefreak/SUN-RGBD", split="train")
row = ds[0]
rgb, depth, labels = row["image"], row["depth_filled"], row["semantic37"]
print(row["scene"], row["sensor"], [o["name"] for o in row["objects_3d"]])

Dataset Sources

Original Paper

SUN RGB-D: A RGB-D Scene Understanding Benchmark Suite

Shuran Song, Samuel P. Lichtenberg, Jianxiong Xiao

IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.

Citation

The authors ask that you cite SUN RGB-D and the datasets it builds on:

@inproceedings{song2015sun,
  title     = {SUN RGB-D: A RGB-D Scene Understanding Benchmark Suite},
  author    = {Song, Shuran and Lichtenberg, Samuel P. and Xiao, Jianxiong},
  booktitle = {CVPR},
  year      = {2015}
}
@inproceedings{silberman2012indoor,
  title     = {Indoor Segmentation and Support Inference from RGBD Images},
  author    = {Silberman, Nathan and Hoiem, Derek and Kohli, Pushmeet and Fergus, Rob},
  booktitle = {ECCV},
  year      = {2012}
}
@inproceedings{janoch2011category,
  title     = {A Category-Level 3-D Object Dataset: Putting the Kinect to Work},
  author    = {Janoch, Allison and Karayev, Sergey and Jia, Yangqing and Barron, Jonathan T. and Fritz, Mario and Saenko, Kate and Darrell, Trevor},
  booktitle = {ICCV Workshop on Consumer Depth Cameras for Computer Vision},
  year      = {2011}
}
@inproceedings{xiao2013sun3d,
  title     = {SUN3D: A Database of Big Spaces Reconstructed using SfM and Object Labels},
  author    = {Xiao, Jianxiong and Owens, Andrew and Torralba, Antonio},
  booktitle = {ICCV},
  year      = {2013}
}

License

The original authors have not specified a license for SUN RGB-D; the project page only requires citation of the papers above. This repository therefore uses license: other and redistributes the data unofficially for research purposes, with attribution. Users are responsible for ensuring their use complies with any terms the original authors may set. Rights holders can request removal by opening a discussion on this repository.