Messy Kitchens - Before & After LiDAR Dataset (Free Sample)
Simulation & Digital Twin Data
Tags and Keywords
Free
About
Real kitchens. Real clutter. Data for your next model.
Download a free sample dataset of 20 U.S. kitchens, captured before and after tidying. Explore photographs, 95 LiDAR scan passes, depth data, walkthrough videos, object annotations, and generated 3D reconstructions for robotics, scene understanding, and embodied AI.
Harvard’s Computational Robotics Group used CollectData kitchen captures to build a robot simulation environment with SceneAgent.
Need more environments, different tasks, or an exclusive collection? CollectData can scope new data collection around your requirements. Use this sample to explore our capture quality and delivery format, then contact us through Opendatabay to discuss a paid pilot or larger project.
This dataset is free for research and commercial AI development under the CollectData Sample Data License 1.1. Privacy, attribution, and source-data redistribution restrictions apply.
Data Product Features
- Before-and-after captures: Observe the same kitchen in messy and tidied states.
- Photographs: Wide views, countertops, and closeups showing household objects and clutter.
- LiDAR and depth: 95 scan passes, with meshes, depth maps, confidence maps, and calibrated camera records where available.
- Object annotations: Contributor-supplied labels and rotated bounding boxes linked to specific images.
- Walkthrough videos: Silent videos with timestamped, machine-transcribed narration labels where available.
- Generated 3D reconstructions: Estimated room and object geometry for reconstruction and simulation experiments.
Available assets vary by kitchen and scan pass.
Distribution
The free full download includes all 20 kitchens. A smaller one-kitchen preview is available for quickly testing the files and loading workflow.
Files are organized by kitchen, before/after state, and scan recording.
- Data Volume: 20 kitchens and 95 scan passes.
- Asset Index:
assets.csvcontains 350 records and 5 columns. - Scan Index:
scan-passes.csvcontains 95 records and 8 columns. - Download Size: Approximately 18 GB compressed for the full dataset. A smaller one-kitchen example is also available.
- Formats: JPEG/HEIC images, PNG depth and confidence maps, MOV videos, PLY/GLB geometry, JSON annotations and metadata, and CSV indexes.
- Documentation: Includes a README, file checksums, and a Python example for loading images, depth, camera records, meshes, and annotations.
Usage
- Scene Understanding: Explore object recognition, clutter interpretation, and spatial relationships in real homes.
- Goal-State Reasoning: Compare messy and tidied scenes to study desired outcomes for household tasks.
- 3D Reconstruction: Combine images, depth, camera poses, and meshes to reconstruct indoor environments.
- Robotics and Simulation: Develop perception pipelines and prototype simulation environments using household geometry.
- Multimodal AI: Connect visual observations, spatial structure, and descriptive annotations.
Coverage
- Geographic Coverage: Residential kitchens in the United States.
- Environment Coverage: Everyday household layouts, objects, surfaces, and clutter, captured before and after tidying.
- Time Range: Exact collection start and end dates are not specified in the release documentation. Individual capture timestamps remain in technical metadata where available. This listing describes release v1.8.0 (2026).
- Demographics: No demographic labels or claims of population representativeness.
License
CUSTOM — CollectData Sample Data License 1.1
Research and commercial AI development are permitted under the full license:
AI Training Rights
Subject to the license, recipients may:
- Train, fine-tune, evaluate, and deploy AI and machine learning models, including commercial models.
- Use the data for research, reconstruction, simulation, and product development.
- Publish and commercialize permitted outputs, including trained models, model weights, generated reconstructions, and research findings.
Rights are worldwide, non-exclusive, and royalty-free, with no scheduled expiry, subject to the license’s termination and affected-data provisions.
Restrictions:
- Source dataset files may not be resold, publicly reposted, or redistributed without permission, except for authorized internal and service-provider use.
- Public display of original source photographs, footage, or audio requires separate permission.
- Publicly shared results must credit CollectData and protect contributors’ privacy.
- Re-identification is prohibited. Applicable privacy, security, and withdrawal requirements must be followed.
The full license controls if this summary differs from its terms.
Who Can Use It
- Robotics and Embodied AI Teams: Develop perception, spatial reasoning, and household-task research.
- Computer Vision Researchers: Explore object recognition, scene changes, and reconstruction.
- AI Labs and Product Teams: Evaluate data pipelines and prototype models before commissioning larger collections.
Data Dictionary
Asset index:
assets.csv| Column Name | Data Type | Description | Possible Values/Notes |
|---|---|---|---|
kitchen | Integer | Kitchen identifier | 1–20 |
asset | String | Asset category | Examples: photo:wide, lidar-scan, walkthrough-video, generated-scene |
state | String | Before/after state | before, after, or blank for assets without a state |
raw_files | String | Paths to capture files | Space-separated paths relative to the dataset root |
derived_path | String | Path to a processed asset | Relative path; may be blank |
Scan index:
scan-passes.csv| Column Name | Data Type | Description | Possible Values/Notes |
|---|---|---|---|
kitchen | Integer | Kitchen identifier | 1–20 |
state | String | Before/after state | before or after |
pass | Integer | Recording number within a kitchen and state | Gaps indicate recordings not included |
video | String | Path to the scan video | Relative path, where available |
frames | String | Path to camera-record JSON | Contains camera calibration and pose information |
keyframes | String | Path to the calibrated keyframe archive | RGB images, depth, confidence, and camera records where available |
mesh | String | Path to the scan mesh | Relative path, where available |
is_primary | Boolean | Whether the recording is selected as the primary scan for browsing | true or false |
Detailed camera, coordinate, and annotation conventions are documented in the included README.
Additional Notes
Scans may contain holes or alignment errors. Generated reconstructions and object placements are estimates; annotations and transcriptions may contain errors.
Privacy treatments include obscuring identifying details, removing location metadata, and stripping video audio. The data is not guaranteed anonymous.
Commission a Custom Collection
Tell us what your model needs: target tasks, environments, capture modalities, data volume, annotations, and delivery requirements.
We can scope additional household captures, different environments or activities, and new collections with agreed exclusivity. Start with a focused paid pilot, review the results, and expand from there.
Contact CollectData through Opendatabay to discuss your collection brief.
Listing Stats
VIEWS
3
DELIVERY
INSTANT DOWNLOAD
LISTED
22/10/2026
UPDATED
23/10/2026
REGION
NORTH AMERICA
TRUST
5 / 5
Loading...
Free
Download Dataset in ZIP Format
Recommended Datasets
Loading recommendations...