pretty_name: data_02
license: other
license_name: aria-digital-twin-dataset-license
license_link: https://www.projectaria.com/datasets/adt/license/
viewer: false
tags:
- computer-vision
- rgbd
- 3d-point-tracking
- raw-data
extra_gated_heading: Request access to data_02
extra_gated_description: >-
Access requests are reviewed manually by the repository owner. Download access
is granted only after approval.
extra_gated_button_content: Submit access request
data_02
This repository contains 152 source recordings selected from the Aria Digital Twin dataset for training with WorldTrack-ADT and TAPVid-3D ADT minival kept separate.
Contents
The repository preserves 608 original files totaling 1,241,197,736,053 bytes (approximately 1.24 TB). Each recording contains four files: the main VRS recording, main ground-truth ZIP, depth ZIP, and instance-segmentation ZIP. All files were checked against the upstream SHA-1 checksums. No extraction, recompression, or training preprocessing has been applied.
| Group | Files | Bytes |
|---|---|---|
| main_vrs | 152 | 277,649,046,732 |
| main_groundtruth | 152 | 7,454,077,849 |
| depth | 152 | 891,424,672,886 |
| segmentation | 152 | 64,669,938,586 |
Access
This is a public repository with manual access approval. Submit an access request on this page and wait for the owner's approval before downloading. Authenticate with your approved Hugging Face account when accessing the files.
The data originates from Aria Digital Twin. The original dataset license applies. Access approval for this repository does not replace the upstream license requirements.
Organization
raw/<sequence_id>/: four upstream files per source recording, with original names.metadata/files.jsonl: relative paths, modality groups, file sizes, and upstream SHA-1 checksums.metadata/sequences.jsonl: the files belonging to each training recording.metadata/provenance.json: dataset source, version reference, and selection policy.splits/train_sequences.txt: the 152 included source recording identifiers.splits/excluded_*.txt: excluded evaluation recordings, synchronized views, and sample duplicates.
Evaluation isolation
The selection excludes the union of 45 WorldTrack-ADT source recordings and 45 TAPVid-3D ADT minival source recordings. Their intersection contains 11 recordings, giving 79 excluded evaluation sources. Four additional recordings from synchronized devices and one sample duplicate are also excluded: 236 - 79 - 4 - 1 = 152 training recordings.
WorldTrack exclusions follow Gen4P Appendix B.3 and Table A2. TAPVid-3D minival identifiers and the v1-to-v2 mapping follow google-deepmind/tapnet at commit 730cda1c730877cfedbe01bf87fb1cadb78a565d.
The isolation scope covers source recordings and synchronized device views for these two evaluation sets. Other TAPVid-3D splits and scene-level independence require separate auditing. These are raw recordings rather than precomputed point-track training examples.
Reading the data
VRS files require Project Aria Tools. Refer to the official ADT data format for extraction, expected filenames, calibration, trajectories, and instance metadata.
Attribution
Please cite the original Aria Digital Twin paper when using this data. This repository is a selected subset of the upstream dataset.