tianxia2/projseg-adt-seq141-subset

ADT seq141 Subset — Projection-Assisted Segmentation Reproduction

A 150-frame slice of the Aria Digital Twin (ADT) sequence Apartment_release_clean_seq141_M1292, packaged so that a projection-assisted gaze-prompted segmentation pipeline can be reproduced without downloading the ~10 GB raw ADT recording.

Companion checkpoints: tianxia2/projseg-checkpoints.

Contents

test/Apartment_release_clean_seq141_M1292/, frames 0–149:

File	Size	Extracts to	What
rgb.tar	396 MB	rgb/frame_%06d.png	RGB, 1408×1408, lossless PNG
depth.tar	190 MB	depth/frame_%06d.npz	Metric depth, 1408×1408
gaze.tar	160 KB	gaze/frame_%06d.json	Gaze pixel + timestamp
segmentation.tar	5.5 MB	segmentation/%06d.npz	Instance segmentation
semantic.tar	5.5 MB	semantic/%06d.png	Semantic labels
aria_trajectory.csv	920 KB	—	Device poses (tx,ty,tz,qx,qy,qz,qw) for the whole sequence
calibration.json	1 KB	—	RGB camera intrinsics (KB8) and T_rgb_device, extracted from the original VRS
metadata.json	44 KB	—	Frame index restricted to this subset

Total ~599 MB. Filenames keep their original frame indices, so frame_000000.png here is frame_000000.png in the full sequence.

The per-frame files are shipped as tar archives, one per modality. Publishing loose per-frame files instead makes a plain snapshot_download exceed the Hugging Face API rate limit (1000 requests / 5 min) partway through. download_assets.py in the reproduction artifact downloads and extracts them in one step.

calibration.json and aria_trajectory.csv are what make the raw ADT recording unnecessary: upstream, the camera model was read from the 1.8 GB VRS via projectaria_tools and the poses from the raw sequence directory.

Usage

huggingface-cli download tianxia2/projseg-adt-seq141-subset \
  --repo-type dataset --local-dir data/processed_adt

cd data/processed_adt/test/Apartment_release_clean_seq141_M1292
for f in *.tar; do tar -xf "$f" && rm "$f"; done

Then follow the reproduction artifact README, which uses this as data/processed_adt/test/Apartment_release_clean_seq141_M1292.

Provenance and license

Derived from the Aria Digital Twin dataset by Meta Platforms, Inc. The underlying data remains subject to its original license and terms of use — see the ADT dataset page. This repository redistributes a frame-limited, format-converted subset purely as a reproducibility convenience. If you use it, cite the ADT dataset and follow its terms; for full sequences or any other use, obtain the data from the official source.

@inproceedings{pan2023aria,
  title     = {{Aria Digital Twin}: A New Benchmark Dataset for Egocentric 3D Machine Perception},
  author    = {Pan, Xiaqing and Charron, Nicholas and Yang, Yongqian and Peters, Scott and
               Whelan, Thomas and Kong, Chen and Parkhi, Omkar and Newcombe, Richard and Ren, Yuheng Carl},
  booktitle = {ICCV},
  year      = {2023}
}