ScanNet is an RGB-D video dataset containing 2.5 million views in more than 1500 scans, annotated with 3D camera poses, surface reconstructions, and instance-level semantic segmentations.
If you would like to download the ScanNet data, please fill out an agreement to the ScanNet Terms of Use and send it to us at [email protected].
If you have not received a response within a week, it is likely that your email is bouncing - please check this before sending repeat requests.
Please check the changelog for updates to the data release.
The data in ScanNet is organized by RGB-D sequence. Each sequence is stored under a directory with named scene<spaceId>_<scanId>
, or scene%04d_%02d
, where each space corresponds to a unique location (0-indexed). The raw data captured during scanning, camera poses and surface mesh reconstructions, and annotation metadata are all stored together for the given sequence. The directory has the following structure:
<scanId>
|-- <scanId>.sens
RGB-D sensor stream containing color frames, depth frames, camera poses and other data
|-- <scanId>_vh_clean.ply
High quality reconstructed mesh
|-- <scanId>_vh_clean_2.ply
Cleaned and decimated mesh for semantic annotations
|-- <scanId>_vh_clean_2.0.010000.segs.json
Over-segmentation of annotation mesh
|-- <scanId>.aggregation.json, <scanId>_vh_clean.aggregation.json
Aggregated instance-level semantic annotations on lo-res, hi-res meshes, respectively
|-- <scanId>_vh_clean_2.0.010000.segs.json, <scanId>_vh_clean.segs.json
Over-segmentation of lo-res, hi-res meshes, respectively (referenced by aggregated semantic annotations)
|-- <scanId>_vh_clean_2.labels.ply
Visualization of aggregated semantic segmentation; colored by nyu40 labels (see img/legend; ply property 'label' denotes the ScanNet label id)
|-- <scanId>_2d-label.zip
Raw 2d projections of aggregated annotation labels as 16-bit pngs with ScanNet label ids
|-- <scanId>_2d-instance.zip
Raw 2d projections of aggregated annotation instances as 8-bit pngs
|-- <scanId>_2d-label-filt.zip
Filtered 2d projections of aggregated annotation labels as 16-bit pngs with ScanNet label ids
|-- <scanId>_2d-instance-filt.zip
Filtered 2d projections of aggregated annotation instances as 8-bit pngs
The following are overviews of the data formats used in ScanNet:
Reconstructed surface mesh file (*.ply
):
Binary PLY format mesh with +Z axis in upright orientation.
RGB-D sensor stream (*.sens
):
Compressed binary format with per-frame color, depth, camera pose and other data. See ScanNet C++ Toolkit for more information and parsing code. See SensReader/python for a very basic python data exporter.
Surface mesh segmentation file (*.segs.json
):
{
"params": { // segmentation parameters
"kThresh": "0.0001",
"segMinVerts": "20",
"minPoints": "750",
"maxPoints": "30000",
"thinThresh": "0.05",
"flatThresh": "0.001",
"minLength": "0.02",
"maxLength": "1"
},
"sceneId": "...", // id of segmented scene
"segIndices": [1,1,1,1,3,3,15,15,15,15], // per-vertex index of mesh segment
}
Aggregated semantic annotation file (*.aggregation.json
):
{
"sceneId": "...", // id of annotated scene
"appId": "...", // id + version of the tool used to create the annotation
"segGroups": [
{
"id": 0,
"objectId": 0,
"segments": [1,4,3],
"label": "couch"
},
],
"segmentsFile": "..." // id of the *.segs.json segmentation file referenced
}
BenchmarkScripts/util_3d.py gives examples to parsing the semantic instance information from the *.segs.json
, *.aggregation.json
, and *_vh_clean_2.ply
mesh file, with example semantic segmentation visualization in BenchmarkScripts/3d_helpers/visualize_labels_on_mesh.py.
2d annotation projections (*_2d-label.zip
, *_2d-instance.zip
, *_2d-label-filt.zip
, *_2d-instance-filt.zip
):
Projection of 3d aggregated annotation of a scan into its RGB-D frames, according to the computed camera trajectory.
Tools for working with ScanNet data. SensReader loads the ScanNet .sens
data of compressed RGB-D frames, camera intrinsics and extrinsics, and IMU data.
Code for estimating camera parameters and depth undistortion. Required to compute sensor calibration files which are used by the pipeline server to undistort depth. See CameraParameterEstimation for details.
Mesh supersegment computation code which we use to preprocess meshes and prepare for semantic annotation. Refer to Segmentator directory for building and using code.
ScanNet uses the BundleFusion code for reconstruction. Please refer to the BundleFusion repository at https://github.com/niessner/BundleFusion . If you use BundleFusion, please cite the original paper:
@article{dai2017bundlefusion,
title={BundleFusion: Real-time Globally Consistent 3D Reconstruction using On-the-fly Surface Re-integration},
author={Dai, Angela and Nie{\ss}ner, Matthias and Zoll{\"o}fer, Michael and Izadi, Shahram and Theobalt, Christian},
journal={ACM Transactions on Graphics 2017 (TOG)},
year={2017}
}
ScannerApp is designed for easy capture of RGB-D sequences using an iPad with attached Structure.io sensor.
Server contains the server code that receives RGB-D sequences from iPads running the Scanner app.
WebUI contains the web-based data management UI used for providing an overview of available scan data and controlling the processing and annotation pipeline.
Code and documentation for the ScanNet semantic annotation web-based interfaces is provided as part of the SSTK library. Please refer to https://github.com/smartscenes/sstk/wiki/Scan-Annotation-Pipeline for an overview.
We provide code for several scene understanding benchmarks on ScanNet:
- 3D object classification
- 3D object retrieval
- Semantic voxel labeling
Train/test splits are given at Tasks/Benchmark.
Label mappings and trained models can be downloaded with the ScanNet data release.
See Tasks.
The label mapping file (scannet-labels.combined.tsv
) in the ScanNet task data release contains mappings from the labels provided in the ScanNet annotations (id
) to the object category sets of NYUv2, ModelNet, ShapeNet, and WordNet synsets. Download with along with the task data (--task_data
) or by itself (--label_map
).
If you use the ScanNet data or code please cite:
@inproceedings{dai2017scannet,
title={ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes},
author={Dai, Angela and Chang, Angel X. and Savva, Manolis and Halber, Maciej and Funkhouser, Thomas and Nie{\ss}ner, Matthias},
booktitle = {Proc. Computer Vision and Pattern Recognition (CVPR), IEEE},
year = {2017}
}
If you have any questions, please contact us at [email protected]
The data is released under the ScanNet Terms of Use, and the code is released under the MIT license.
Copyright (c) 2017