Access is granted immediately after you submit this form. Your Hugging Face username and email are shared with CollectData for access administration and dataset notices.

I have read and accept the CollectData Sample Data License 1.1 linked above. If accepting for an organization, I have authority to bind it.

Tidy Up: U.S. kitchens before and after cleaning

Real kitchens. Real clutter. More than photos.

Explore 20 U.S. kitchens before and after cleaning, with photos, LiDAR scans, depth data, object annotations, and generated 3D reconstructions. Study how everyday spaces change—and connect what’s visible in an image to the geometry of the room.

A practical sample for exploring scene understanding, 3D reconstruction, and robotics perception. Browse individual kitchens or download the dataset to start experimenting. The sample includes 95 scan passes; available assets vary by kitchen.

Browse a kitchen below, or download everything. There is also a one-kitchen ZIP and ZIP checksums.

Browse the kitchens

LiDAR shows the measured scan mesh; Model shows the generated reconstruction. Use the links below each preview to browse the files.

Kitchen	After	LiDAR	Model
1		Before · After	Model files
2		Before · After	Model files
3		Before · After	Model files
4		Before · After	Model files
5		Before · After	Model files
6		Before · After	Model files
7		Before · After	Model files
8		Before · After	Model files
9		Before · After	Model files
10		Before · After	Model files
11		Before · After	Model files
12		Before · After	Model files
13		Before · After	Model files
14		Before · After	Model files
15	(scan-frame stand-in)	Before · After	Model files
16		Before · After	Model files
17		Before · After	Model files
18		Before · After	Model files
19		Before · After	Model files
20		Before · After	Model files

Find and use the files

Each kitchen has a numbered folder. Photos and scans are grouped by before/after state; annotations and walkthroughs belong to before. Generated reconstructions remain at kitchen level.

kitchens/01/
  before/
    photos/
    annotations/
    scan/           # primary scan files
      pass-NN/      # additional recordings, where available
    walkthrough.mov
  after/
    photos/
    scan/           # primary scan files
      pass-NN/      # additional recordings, where available
  reconstruction/
preview-assets/      # gallery thumbnails only

assets.csv lists kitchen, asset, state, capture files (raw_files) and processed file (derived_path). Paths and annotation imagePath values are relative to the repository root. Reconstruction references are relative to their JSON file. Registered scans include alignment.json with repository-relative mesh paths.

To download just one kitchen and run the included example:

hf download CollectDataIO/tidy-up-kitchens --repo-type dataset --local-dir tidy-up \
  --include 'kitchens/04/**' assets.csv quickstart.py LICENSE.txt
python -m pip install numpy Pillow pillow-heif open3d
python tidy-up/quickstart.py tidy-up --kitchen '4' --output /tmp/tidy-up-example

Sign in with hf auth login first if access is gated. The example loads RGB, depth, confidence, camera poses, a measured mesh and region annotations. datasets.load_dataset loads the asset index; it does not assemble the 3D scenes.

The optional ZIPs use this same before/after layout. Follow the README inside a ZIP when using it. The root manifest.json checksums the browsable files. Updates can be reproduced by pinning a Hugging Face commit. Keep captures from the same kitchen together when splitting data for evaluation; no train/test split is prescribed.

Technical reference and limitations

Data	How to use it
Task photos	Full-resolution JPEGs. A photo marked `stand-in` is a scan frame.
Scan keyframes	`keyframes.zip` contains matching HEIC RGB, `.depth.png`, `.conf.png` and JSON camera records.
Depth / confidence	Depth is millimetres; divide by 1000 for metres. Confidence is ARKit 0–2, not a probability.
Camera records	Intrinsics describe the RGB resolution; scale x/y to the depth resolution when back-projecting. Poses are camera-to-world, metres, Y-up. `sourceFrameNumber` joins keyframes to `frames.json`.
Measured meshes	ARKit meshes use metres and Y-up. `mesh-aligned.ply` is registered to the messy scan frame; independent scans otherwise have separate local frames.
Generated meshes	Rooms use Z-up; object assets marked `up: "y"` need a −90° X rotation. Placements are estimates. Photogrammetry meshes are reconstructed from photos.
Region annotations	Annotation JSON gives an explicit `imagePath`, pixel dimensions and rotated boxes (`centerX`, `centerY`, `width`, `height`, `rotationRadians`, `label`). Use that image, not a rotated preview; scale coordinates if resizing.
Walkthroughs	Videos are silent. Narration labels are machine-transcribed text with timestamps and may contain errors.

Limitations

Scans can have holes and registration errors. Generated reconstructions and object positions are approximate. Labels are contributor annotations, not a validated benchmark. No train/test split is prescribed.

Faces and private details have been obscured, location metadata removed and video audio stripped. Some assets are omitted; masked areas contain no usable appearance information. The data is not guaranteed anonymous. Capture times may remain in technical metadata.

Kitchen-specific limitations are listed below. File availability is shown directly in the asset index rather than in a second summary table.

Kitchen 2: primary scans have no video; additional passes include video. The generated room has neutral colors.
Kitchen 3: clean scan and walkthrough cover a side table rather than the full kitchen.
Kitchen 6: counterMessy photo is a scan-frame stand-in.
Kitchen 15: wideClean photo is a scan-frame stand-in.

License and citation

CollectData Sample Data License 1.1 allows research, commercial development and publishing generated models/reconstructions, subject to privacy and attribution requirements. Credit CollectData on your project page or in accompanying documentation. The license governs source-media use and redistribution.

Cite CollectData, Tidy Up: kitchen sample, v1.7.0 (2026), with this repository URL and the commit used. Contact info@collectdata.io for support or privacy concerns.

v1.7.0 simplifies documentation and gives region annotations explicit image paths. This browsable edition groups its files by kitchen; captured media, calibration, geometry and object placements are unchanged.

Separate scan passes capture different views. scan-passes.csv lists every available pass and its files in capture order. Each recording is provided once, with no separate preview video.

Passes have independent local coordinate systems unless an alignment file is provided.

Scan recordings retain their captured timing. Kitchen 18’s before scan is limited to the first 105 seconds, with matching camera records and keyframes; its walkthrough is limited to the first 65 seconds. Playback rotation metadata makes consistently sideways recordings upright; for calibrated pixel access, disable automatic video rotation. Some captures change orientation during recording.

Some passes have video and camera records but no RGB-D keyframes; scan-passes.csv lists the available files.

Downloads last month: 42

Total file size:

37.4 GB