This dataset is derived from 3D-FRONT / 3D-FUTURE (Alibaba / Tao Bao) and may be used for scientific research purposes only. By requesting access you agree to the original 3D-FRONT Data Sets Use License Agreement (included as 3D-FRONT-ToU.pdf) and accept that you will pass the same terms on to anyone you share this data with.

Indoor Scene Renders from 3D-FRONT / 3D-FUTURE

20,240 photo-realistic indoor scene renders with instance segmentation, 6DoF object poses, camera intrinsics and depth — rendered from the 3D-FRONT scene layouts and 3D-FUTURE furniture models.

This is not a copy of the original 3D-FUTURE render set. It is a separate render set built from the same assets. See Differences from the original below.

Why this exists

The 3D-FUTURE technical report describes 20,240 rendered indoor images, but those images were never part of any public release — the mirrors that carry 3D-FRONT and 3D-FUTURE contain scenes, furniture and textures only. The official Tianchi distribution pages now return 404 and the listed contact address does not answer, so the images cannot be obtained from the source either.

This repository provides a render set of the same size, produced with a documented and reproducible pipeline.

What is in here

Folder	Contents
`renders/`	20,240 rooms in 102 WebDataset tar shards (~148 MB each, 15.1 GB total), plus `shard_index.json` and `SHA256SUMS`.
`defects/`	`3d-front-defects.json` — defects we found in the original data, as identifiers and measurements.
`code/`	The full pipeline: selection, rendering, verification, repair, packing.
`docs/`	The original terms of use, our Korean translation, and the working plan.

Reading the shards

Each room contributes four members sharing one key, so any WebDataset-style loader groups them automatically:

<scene_id>_<room_id>.rgb.png
<scene_id>_<room_id>.depth.png
<scene_id>_<room_id>.instance.png
<scene_id>_<room_id>.json

import webdataset as wds

dataset = wds.WebDataset("renders/renders-{000..101}.tar").decode()
for sample in dataset:
    rgb = sample["rgb.png"]
    depth = sample["depth.png"]        # 16-bit, millimetres
    instance = sample["instance.png"]  # pixel value = placement_index + 1
    meta = sample["json"]

Plain tarfile works too. shard_index.json maps every room name to the shard holding it, under the key room_to_shard, so a single room can be pulled without downloading everything. SHA256SUMS covers all 102 shards — verify with sha256sum -c SHA256SUMS. Shards are built deterministically: member order is sorted and timestamps are fixed, so repacking the same renders reproduces the same bytes.

Per-room files

Member	Meaning
`<key>.rgb.png`	1024 × 768 colour image
`<key>.depth.png`	16-bit depth. Millimetres. `0` means background or beyond 20 m.
`<key>.instance.png`	Instance segmentation. Pixel value = `placement_index + 1`; `0` is background.
`<key>.json`	Resolution, camera intrinsics, field of view, camera pose (4×4), and per-object 6DoF pose, scale and bounding box.

Coordinate systems differ by field, and the file says so in its own coordinate_note. Object positions and rotations are in the original y-up frame, with quaternions in scalar-last [x, y, z, w] order. Bounding boxes and camera_to_world are in Blender's z-up frame.

What this is not

We do not redistribute 3D-FRONT or 3D-FUTURE. To reproduce the renders you need to obtain those datasets yourself and accept their terms. This repository carries only what we made: the images, our annotations, our defect measurements and our code.

Differences from the original render set

The 3D-FUTURE paper's images and ours are not interchangeable.

	Original	Ours
Images	20,240	20,240
How the count is filled	20,240 designs built on 5,000 designer rooms, one image each	20,240 different rooms, one image each
Camera	Chosen by professional designers	Chosen automatically, with clearance and framing constraints
Renderer	V-Ray 6	Blender Cycles (BlenderProc 2.8.0)
Resolution	Not published	1024 × 768 (our choice)
Field of view	Provided per image, value not published	60° horizontal (our choice)
Instances per image	5.1	5.4

Because the renderer and the camera policy differ, these images are not a substitute for the originals in any pixel-level comparison, and results measured on them should not be placed side by side with numbers from the original set without saying so.

Defects we found in the original data

Published as defects/3d-front-defects.json so anyone can apply the same quality filtering to data they obtain themselves.

Defect	What goes wrong
Furniture meshes authored in centimetres	The model enters the scene at 100–153× its declared size. One children's bed measured 220 m × 122 m × 186 m. Scale values look normal, so filtering on scale does not catch it.
Truncated or missing textures	Three files are truncated — 76–99% of the file is empty, so the object renders almost black, and the file still opens without error. One is missing outright.
`NaN` placement coordinates	Standard JSON has no `NaN`, so annotations carrying it fail to parse in some languages.

The manifest also lists placements whose rotation is not a pure yaw, and rooms that ATISS-style preprocessing discards because of out-of-range scale values.

Reproducing

Each step feeds the next, so the order matters.

python verify_recovery.py 3D-FRONT-raw            # archive integrity
python prepare_dataset.py 3D-FRONT-raw 3D-FRONT-data
python select_rooms.py 3D-FRONT-data --target 20240  --out render_rooms.json
python select_rooms.py 3D-FRONT-data --target 999999 --out all_candidates.json
python census_broken_textures.py                  # must run before rendering
bash run_render_batches.sh 100
python fill_missing_rooms.py --apply              # then re-run the render step
python repair_renders.py --delete                 # then re-run the render step
python pack_shards.py --out renders --rooms-per-shard 200

build_defect_manifest.py regenerates defects/3d-front-defects.json from the same data. On one RTX 5080 the render step took 10.3 hours (1.84 s per room) and produced 15.06 GB.

The renderer is restarted every 100 rooms on purpose: in a single long-lived process the time per room climbs steadily (1.9 s → 4.7 s over 330 rooms). We did not find the cause; restarting caps it, and costs about 5 s per batch.

Licence

Derived from 3D-FRONT / 3D-FUTURE. The original 3D-FRONT Data Sets Use License Agreement (docs/3D-FRONT-ToU.pdf) applies to this work:

Scientific research purposes only. No commercial use of the data, of works derived from it, or of results obtained with it.
Pass the terms on. If you share this data or anything derived from it, the recipient must be bound by the same agreement.
Cite the source. 3D-FRONT and 3D-FUTURE, plus this repository.

We claim no licence over the underlying 3D-FRONT / 3D-FUTURE content and grant none. 3D-FUTURE carries its own separate agreement (dated 2020-03-28) which we were unable to obtain; if you need certainty about the furniture models, consult it.

Citation

@article{fu2021_3dfront,
  title   = {3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics},
  author  = {Fu, Huan and Cai, Bowen and Gao, Lin and Zhang, Ling-Xiao and Wang, Jiaming and Li, Cao and Zeng, Qixun and Sun, Chengyue and Jia, Rongfei and Zhao, Binqiang and Zhang, Hao},
  journal = {arXiv preprint arXiv:2011.09127},
  year    = {2021}
}

@article{fu2021_3dfuture,
  title   = {3D-FUTURE: 3D Furniture Shape with TextURE},
  author  = {Fu, Huan and Jia, Rongfei and Gao, Lin and Gong, Mingming and Zhao, Binqiang and Maybank, Steve and Tao, Dacheng},
  journal = {International Journal of Computer Vision},
  year    = {2021}
}

Downloads last month: 13

Total file size:

15.1 GB

Paper for Spatial1ntelligence/3d-front-indoor-renders

3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics

Paper • 2011.09127 • Published Nov 18, 2020