The Zillow Indoor Dataset (ZInD) provides extensive visual data that covers a real world distribution of unfurnished residential homes. It consists of primary 360º panoramas with annotated room layouts, windows, doors and openings (W/D/O), merged rooms, secondary localized panoramas, and final 2D floor plans. The figure above illustrates the various representations (from left to right beyond capture): Room layout with W/D/O annotations, merged layouts, 3D textured mesh, and final 2D floor plan.
Definitions: Primary panoramas are those selected by annotators as having the "best" views of entire rooms, and are used to generate room layouts. The rest of the panoramas are secondary panoramas, which are provided for denser spatial data; they are localized within room layouts using a semi-automatic approach. An opening is an artificial construct that divides a large room into multiple parts. Note that openings are later processed for removal.
Zillow Indoor Dataset: Annotated Floor Plans With 360º Panoramas and 3D Room Layouts
Steve Cruz*, Will Hutchcroft*, Yuguang Li, Naji Khosravan, Ivaylo Boyadzhiev, Sing Bing Kang
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021
[Paper] [Supplementary Material]
(* Equal contribution)
If you use the ZInD data or code please cite:
@inproceedings{ZInD,
title = {Zillow Indoor Dataset: Annotated Floor Plans With 360º Panoramas and 3D Room Layouts},
author = {Cruz, Steve and Hutchcroft, Will and Li, Yuguang and Khosravan, Naji and Boyadzhiev, Ivaylo and Kang, Sing Bing},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2021},
pages = {2133--2143}
}ZInD is an RGB 360º panoramas dataset containing 67,448 panoramas taken in 1,575 unfurnished residential homes, annotated with 3D room layouts, 2D bounding boxes for W/D/O, merged room layouts, 3D camera poses, and final 2D floor plans.
The stats reported here are slightly different from those reported in the CVPR paper. If you need to cite the stats, please use the stats reported here.
Please refer to data organization for more details.
To capture home interiors at scale, we opted for sparse 360º panorama capture of every room in the home using an off-the-shelf 360º camera (such as Ricoh Theta V or Z1) paired with an iPhone. To do so, photographers across 20 US cities and 12 states were hired to do the capture; they were given specific instructions to ensure uniformity in the capture quality. Please see the FAQ section for more details.
The 3D floor plans were generated from a sparse set of RGB panoramas using our proprietary human-in-the-loop pipeline. Please see the FAQ section for more details.
A one-home sample tour is given in sample_tour.
If you are interested in downloading ZInD, please register an account on the Bridge Platform. During registration you will be required to agree to the Zillow Data Terms of Use. Before you take all these steps, however, we strongly suggest that you review the FAQ section first. Once your request has been approved, you will be notified. Please do not contact us unless it has been over two weeks since the request was made.
For non-commercial, academic institutions outside US and Canada, please select Other as a State/Province.
In Bridge Platform API, you can find the Server Token under API ACCESS Tab.
Set up a conda environment:
conda create -n zind python=3.6
conda activate zind (Mac)
activate zind (Windows)
Install dependency libraries:
pip install -r requirements.txt
Use download_data.py to automatically download the data from the Bridge Platform, there is no need for you to interact with the API directly. Please note that the size of ZInD is about 40GB, so please make sure you have enough disk space before you start the download process.
python download_data.py -s <server_token> -o <output_folder>
The CoVis scores used in the paper can be found here. The field names are self-explanatory.
Statistics for 1,575 homes and 67,448 panoramas. pri = primary, sec = secondary, “# annotator spaces” refers to spaces identified by annotators (which include closets and hallways), and “# rooms” refers to complete room layouts.
| Feature | Total | Avg Per Home |
|---|---|---|
| # panoramas (pri) | 33210 | 21.086 |
| # panoramas (sec) | 34238 | 21.738 |
| # floor plans | 2737 | 1.738 |
| # annotator spaces | 29410 | 18.673 |
| # rooms | 22484 | 14.276 |
| # windows | 19403 | 12.319 |
| # doors | 48759 | 30.958 |
Statistics on different room layout types. Since L-shaped layouts are common, we report that separately from others that are also Manhattan. Those that are non-Manhattan typically have room corner angles of 135 degree.
| Layout Types | Cuboid | Manhattan-L | Manhattan-General | Non-Manhattan |
|---|---|---|---|---|
| # Layouts | 11924 | 3273 | 2715 | 4572 |
| # Corners | 4 | 5 | 6 | 7 | 8 | 9 | 10+ |
|---|---|---|---|---|---|---|---|
| # Layouts | 11999 | 1002 | 3631 | 370 | 1416 | 154 | 3912 |
ZInD is partitioned to train/val/test (0.8 : 0.1 : 0.1). Those splits have similar distributions under the following metrics:
- Layout complexities (cuboid, L-shape, etc.)
- Number of floors
- Number of primary panoramas
- Number of secondary panoramas
- Total area (to ensure that we have good balance between small/large homes)
The recommended train/val/test splits are in zind_partition.json
The partition script is available at code/partition_zind.py
Our dataset and training split has changed since the time of original submission. As such, the published evaluation numbers should be reproduced with the newly released train/test/val split for any future publications.
Run the visualization script
code/visualize_zind_cli.py
python code/visualize_zind_cli.py -i sample_tour -o <output_folder> --visualize-layout --visualize-floor-plan --raw --complete --visible --primary --secondary
More examples about how to run the visualization script can be found here
The visualization results for the sample tour are in render_data.
Structured3D is a large-scale photo-realistic dataset containing 3.5K house designs proposed in a ECCV'20 paper. To extend the compatibility and potential applications of our ZInD dataset, we provided python scripts which convert original ZInD data format to Structure3D format with minimal modifications, and render the converted samples for visualization purposes. The modifications are due to a few extra components (e.g., windows, doors, and openings). In addition, we further provided the modified visualization scripts for visualizing the converted ZInD samples in an interactive manner.
A "s3d" branch is created for this update. Please see the scripts and details here.
The data is released under the ZInD Terms of Use, and the code is released under the Apache License.
We would like to thank the Zillow RMX engineering team for building the annotation infrastructure, and the ZO photographers for capturing the empty homes. We are grateful to Pierre Moulon and Lambert Wixson for discussions on ZInD. In addition, Ethan Wan has been very helpful in processing and validating data in preparation for release. Finally, the Bridge team has also been instrumental in making ZInD release a reality.
- Can ZInD be downloaded free-of-charge?
- What are the conditions for using ZInD?
- How long do I wait after requesting for ZInD download?
- What are the important issues associated with the dataset that we need to be aware of?
- What was the capture protocol?
- What was the annotation pipeline?
- What if I would like to download ZInD for commercial purposes?
Yes, ZInD is free of charge for academic, non-commercial use.
Zillow licenses ZInD for academic use only. ZInD is not licensed for commercial purposes. Here are some examples of what are acceptable and unacceptable uses of ZInD:
- Acceptable use: You are a researcher (e.g., graduate student, post-doc, professor) at a university, and are using ZInD to investigate potentially new ideas on room shape and floor plan generation from panoramas. You publish a paper based on that research work, and use images and derived data from ZInD.
- Possible acceptable use: You work at a non-academic institution, and you are writing a paper on estimating room shapes from images. You use ZInD as one benchmark, and publish a paper to show the effectiveness of your algorithm. None of that work will be used in any product at your company. Here, you will not be able to license ZInD via Zillow Bridge API, but will need to contact us first with a proposal (500 words minimum). You can contact us for approval of your proposal by emailing your proposal to [email protected].
- Unacceptable use: You work at a non-academic institution, and you use data from ZInD to improve algorithms in a product or service.
- Unacceptable use: You are a researcher at a university, and you collaborate with someone who works at a non-academic institution. While your main goal is to publish a paper, the results of your work based on ZInD data are incorporated in a product or service at that company.
Please note that we need to be able to verify your academic credentials, ideally in the form of a website of the institute to which the requestor belongs. The email address should also reflect the institute. If we are unable to verify, we will reject the request. At this time, we are not able to accommodate requests for personal use.
Please note that the approval process is manual. The decision to approve or deny the request may take up to two weeks.
Given the manner in which the panoramas were captured and annotated, there are specific characteristics that needed to taken into consideration. Details can be found in the CVPR 2021 paper. Here we list some notable features/issues:
- Visible geometry will not be available for very small rooms (e.g., closets), where the panorama is captured outside such a room.
- Small amount of annotation errors might be present in all tasks, like room shapes, WDO boundaries and secondary panorama localization.
- There are occasional, pre-redraw, artifacts due to separating the annotation tasks, such as door dimensions not matching exactly at adjacent rooms.
- Once in a while, the IMU data and/or our upright correction algorithm fails, resulting in a panorama that is tilted.
- For complex (non-flat) ceilings, the given height is only an approximation.
- For "visible" geometry, note that partially visible geometries are locally clamped to what is observable, so that the extracted geometry would no longer be Manhattan (assuming the original room layout is).
- "Complete" and "visible" geometry will not extend through doors, but only openings, since we don't collect annotations on whether a door is closed or open.
- The merger (pre-redraw) geometry and the final redraw geometry would not align perfectly due to the final human touch up in redraw to create a polished and globally consistent floor plan.
- There are floorplans with rooms labeled "master bedroom" and "master bathroom". These are deprecated terms; please consider them as "primary bedroom" and "primary bathroom," respectively.
- There are floor plans with no scale (specifically, 'floor_XX' : None), and this is typically caused by issues in calibration.
There are also rare cases of the following:
- Duplicates.
- Incorrectly annotated doors and windows, e.g., the top edge of a door reaches the ceiling.
- Annotators failing to refine poses of secondary panoramas.
- Severely underexposed panoramas due to incorrect merging of multiple exposures.
- Self intersecting geometry due to incorrect annotations.
To enable capturing entire home interiors at scale, we opted for sparse 360º panorama capture of every room in the home using an off-the-shelf 360º camera (such as Ricoh Theta V or Z1) paired with an iPhone. Photographers across 20 US cities were hired to do the capture. To ensure consistency in the capture process, they were given specific instructions, which include:
- Capture every room, including connecting hallways and garages.
- Keep a fixed tripod height within a home.
- Capture a calibration target to allow camera height to be computed.
- Keep interior doors open whenever possible.
- Turn on all lights and turn off fans and TVs.
- Avoid capturing personal information (such as people, photographs of people, pets, and objects).
Please refer to the Supplementary Material for more details.
To generate 3D floor plans from a sparse set of RGB panoramas, we developed a proprietary human-in-the-loop pipeline. Our pipeline starts with automatic pre-processing of panoramas, which includes straightening, room layout estimation, and W/D/O detection. Subsequently, trained annotators are tasked with:
- Selecting, verifying, and correcting primary room layouts and W/D/O features.
- Merging verified primary room layouts to form a draft floor plan.
- Localizing secondary panoramas within the existing, primary layouts.
- Fixing and cleaning up the draft 2D floor plan to generate the final version.
Please refer to the Supplementary Material for more details.
To reduce privacy concerns, we have automatically detected and removed any panoramas that contain:
- People or photographs of people to avoid sharing PI or PII information.
- Significant outdoor view, to avoid sharing street views or neighbouring properties.
Please email us, and we will forward your request to our business representative.
