By requesting access you agree to the EgoCrowd-500H Preview Terms, including that redistribution, model training, benchmarking beyond the paper's evaluation scripts, or production use requires a separate written agreement with Poseidon, and that you will not attempt to re-identify participants, workplaces or devices.

EgoCrowd-500H — preview release

300 stereo egocentric clips · 50 with temporally localised human review · 4 failure families
Companion release for When Crowdsourced Egocentric Video Goes Wrong: Failure Modes, Metadata, and Quality Control (CoRL 2026 workshop) · Paper (PDF) · Interactive explorer

Need a custom egocentric dataset? Talk to Poseidon →

What this is

Crowdsourced egocentric video scales quickly, but more footage is not more usable footage. The paper audits 500 hours / 3,584 recordings of crowdsourced first-person video and organises what goes wrong into four failure families — capture, visual quality, task, and metadata — grounded in 81.1 hours of temporally localised human review. Its central finding is that failures are usually localised: an imperfect recording is rarely imperfect throughout, so recording-level accept/reject discards far more footage than the annotated failure intervals.

This repository is the first public slice of the data behind that study. It contains 300 fixed-length 30 s stereo clips drawn from the corpus with a seeded, stratified sample that deliberately spans the quality spectrum (not just the good footage), together with the metadata each clip carried in the collection pipeline, the automatic QC signals that were used for sampling, and — for 50 clips — frame-accurate human review of idle time and off-task behaviour. The remaining footage described in the paper will be released in the same format.

Highlights

Stereo, full resolution. Every clip is a side-by-side stereo frame at 4000x1200 (two 2000×1200 eyes), ~30 fps (effective 19.0–30.1 fps, variable-frame-rate sources), H.264. Nothing is cropped, re-timed or resampled.
Spans the quality spectrum. Clips were sampled by QC stratum so that likely-acceptable, borderline and likely-rejected footage are all present (see below). Failure evidence is kept, never cropped away.
Temporally localised review. 50 clips carry reviewer-marked idle and off-task spans with free-text descriptions, mapped onto the paper's taxonomy.
Honest QC signals. Automatic CLEF scores are provided as the sampling signal they were, attached to the 30 s source chunk, and are explicitly not a verified label for the clip.
Privacy-first. Every clip was screened on both stereo views for faces and readable text; any clip with a face or with document/screen-like text was excluded from this release and container metadata was stripped. Participant, device and internal identifiers are not present.

At a glance


Clips	300 × 30 s (22.6 GB)
Media	MP4 / H.264 High@L5.1, 4000x1200 side-by-side stereo, ~30 fps (effective 19.0–30.1)
Countries	Mexico (173), Egypt (118), Argentina (9)
Task families	37
Jobs / skill groups	22 / 11
Environments (L1)	11
Human-reviewed clips	50
Clips with blurred faces	0 (clips with detected faces were excluded instead)
License	Poseidon Preview Terms (custom, non-open) — see License and access
Access	Gated — company name, role and company email required

QC strata

Stratum	Clips	Share
Likely acceptable (CLEF above threshold)	210	70.0%
Borderline (within ±0.02 of threshold)	60	20.0%
Likely rejected (CLEF below threshold)	30	10.0%

CLEF is the automatic clip-level quality classifier used in the collection pipeline (two policy versions, thresholds 0.76 and 0.656). Strata are defined on the source chunk's score and were used only to make the sample span the quality range. They are not ground truth.

Environments (L1)

Environment	Clips	Share
Industrial Manufacturing	102	34.0%
Administrative & Office Services	65	21.7%
Healthcare and Pharmacy	53	17.7%
Repair Services	19	6.3%
Unspecified	16	5.3%
Printing and Design	16	5.3%
Food and Beverage	10	3.3%
Construction and Hardware	8	2.7%
Cleaning and Sanitation	7	2.3%
Hospitality	2	0.7%
Beauty and Personal Care	2	0.7%

Countries

Country	Clips	Share
Mexico	173	57.7%
Egypt	118	39.3%
Argentina	9	3.0%

Failure families

Family	Definition	In this release
Capture	Failures of the recording process: camera misalignment, obstruction, wrong mounting.	Visible in footage; not yet labelled per clip
Visual quality	Reduced observability: blur, exposure, poor lighting.	Visible in footage; not yet labelled per clip
Task	Deviations from the intended activity: idle time, off-task behaviour, staged or repeated actions, no visible task.	24 clips with reviewer-localised spans
Metadata	Content inconsistent with its task description or other metadata.	Reviewer-assigned task id vs. supplied task description, where available

Privacy annotations are tracked separately from the four families (pii_screen).

Human review (50 clips)

Reviewers followed an eight-gate protocol (capture → visual quality → task identification → behavioural authenticity → skill → job → environment → privacy) and marked idle spans (both hands still or nothing task-related in hand for > 2 s) and non-task spans (phone use, conversation, walking without work, smoking, eating/drinking, unrelated object, screen unreadable, other), plus staged / repeated-without-progress behaviour. Each span has a start, end and a one-line description. We also publish a mechanically derived verdict (verdict_derived) obtained by applying the protocol's G3 rules to those spans:

Derived verdict	Clips
accept	26
reject	16
accept_with_crop	8

The derived verdict is a convenience for analysis, not a reviewer decision. Reviewer accounts are pseudonymised (reviewer_01 …).

Full stereo recordings (`stereo/`)

Alongside the fixed-length clips, the repository carries 7 complete stereo recordings (1.5 GB) from the paper's two stereo QA cohorts — separate left/right eye videos plus the synchronised sensor streams that were captured with them.


Recordings	7 (CN (4), KR (3))
Headsets	VST stereo headset (4), Eye2 stereo headset (2), Dual-EGO stereo headset (1)
Per-eye resolution	2328x1748 (4), 1920x1200 (2), 1600x1200 (1)
Sensor streams	IMU (all), per-frame capture timestamps (all), audio (4), device pose (4)
Layout	`stereo/<record_id>/clip.json`, `video/ego_left.mp4`, `video/ego_right.mp4`, `video/.timestamps.parquet`, `sensors/.parquet`, `sensors/audio.wav`

Video streams are byte-identical to the source captures (lossless remux; container tags stripped). clip.json follows the Poseidon ego release schema: task, environment, cameras (with per-file SHA256), streams, clocks, and ingest notes; vendor paths and source URLs are removed.

Why only 7? The stereo cohorts total 935 recordings (81 h). Each was screened at 1 fps with two face detectors and OCR; 928 showed a face or identifier-like text and were withheld — these are 5-minute recordings in shared workplaces, so bystander faces are nearly universal. The withheld recordings exist and are described in the paper; they will be added only after a redaction policy is agreed. Per-recording screening statistics are in stereo_metadata/recordings.jsonl (pii_screen).

Repository layout

data/clips/ec_0001.mp4 … ec_0300.mp4   full-resolution stereo clips (HF `videofolder` layout)
data/metadata.jsonl                       one row per clip (file_name relative to data/)
previews/ec_XXXX.mp4                      1280-px-wide previews for browsing
posters/ec_XXXX.jpg                       left-eye poster frame
metadata/clips.parquet, clips.jsonl       the same rows with repo-root-relative paths
taxonomy.json                             failure families, review protocol values, stratum definitions
schema.md                                 field-by-field description of metadata.jsonl
eval/                                     scripts: validate checksums/media, reproduce the sampling strata, summarise review spans
SHA256SUMS.txt, dataset_info.json
stereo/<record_id>/                       full multi-sensor stereo recordings (see section above)
stereo_previews/, stereo_posters/          15 s side-by-side previews and posters of those recordings
stereo_metadata/                          recordings.{jsonl,parquet}, files.parquet, SHA256SUMS-stereo.txt, stereo_info.json

Usage

This dataset is gated. Request access on this page, then log in with huggingface-cli login (or pass token=...) before downloading.

from huggingface_hub import snapshot_download
import pandas as pd

root = snapshot_download("psdn-ai/EgoCrowd-500H", repo_type="dataset")          # gated: log in or pass token=...
clips = pd.read_json(f"{root}/metadata/clips.jsonl", lines=True)
reviewed = clips[clips.has_human_review]
print(clips.qc_stratum_group.value_counts())

Split a stereo frame into eyes with ffmpeg: ffmpeg -i ec_0001.mp4 -filter_complex "[0:v]crop=iw/2:ih:0:0[left];[0:v]crop=iw/2:ih:iw/2:0[right]" -map "[left]" left.mp4 -map "[right]" right.mp4.

Validate your download: python eval/validate.py --root .

Privacy and redaction

Both stereo views of every candidate clip were screened on every third frame with two face detectors (InsightFace SCRFD and OpenCV YuNet) and every 30th frame with EasyOCR (en/es).
Detections were reviewed visually; any clip with a visible face was excluded from the release. Clips whose readable text looked like documents, screens or numeric identifiers were also excluded. Residual signage/brand tokens that were read are listed per clip in pii_screen.visible_text_tokens.
Container metadata was stripped; participant, device, camera-serial, operator and internal path identifiers are not published. Public ids (ec_XXXX) are not derivable from internal ids.
Footage was recorded by participants enrolled in Poseidon's egocentric data-collection program, in their own workplaces. If you believe something should be removed, contact hello@psdn.ai.

Limitations

The QC scores are a sampling signal. They describe the source chunk and the two policy versions are not directly comparable; see qc_note per row.
Task labels (task_description, job, skill_group, task_difficulty_supplied) are the labels supplied at collection time; for ~10% of clips they were inferred from sibling recordings (metadata_provenance).
Human review covers 50 clips and one review pass each; inter-annotator agreement was not measured on this subset.
Capture and visual-quality failures are present in the footage but not yet labelled per clip in this preview.

About Poseidon

Poseidon builds the training data for physical AI. Machines learn physical skill the way apprentices always have, by watching skilled people work, and that footage has to be created on purpose and to specification. Poseidon runs that supply chain: it recruits skilled people and real workplaces around the world, records their work to exact lab specifications, and refines the footage into model-ready datasets.

Learn more about Poseidon →

License and access

This preview is not released under an open license. It is provided under Poseidon's custom EgoCrowd-500H Preview Terms (license: other, license_name: poseidon-sample-preview) — see TERMS.md. In short:

Permitted: reviewing the dataset's structure, media quality and annotation format, and reproducing the analyses in the paper.
Requires a separate written agreement with Poseidon: redistribution, model training, benchmarking beyond the paper's evaluation scripts, and production use.
Prohibited: attempting to re-identify participants, workplaces or devices, or to reverse any redaction.

Access is gated: you will be asked for your company name, role and company email, and to accept the Preview Terms.

For the full corpus, commercial licensing, or custom coverage, contact Poseidon:

Email hello@psdn.ai

Not what you're looking for? Request a custom dataset →

Citation

If you use this dataset, please cite both the paper (for the failure taxonomy, review protocol and findings) and the dataset release (for the specific data and version you used).

Paper

@inproceedings{nagpal2026egocrowd,
  title     = {When Crowdsourced Egocentric Video Goes Wrong: Failure Modes, Metadata, and Quality Control},
  author    = {Rashmi Nagpal and Lefei Liu and Jongwon Park and Sandeep P. Chinchali},
  booktitle = {Conference on Robot Learning (CoRL) Workshops},
  year      = {2026},
  url       = {https://openreview.net/pdf?id=rLGRN1Uoh2}
}

Dataset

@misc{poseidonai2026egocrowd500h,
  title        = {{EgoCrowd-500H (preview release)}},
  author       = {Rashmi Nagpal and Lefei Liu and Jongwon Park and Sandeep P. Chinchali},
  year         = {2026},
  version      = {0.1.0-preview},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/psdn-ai/EgoCrowd-500H},
  howpublished = {\url{https://huggingface.co/datasets/psdn-ai/EgoCrowd-500H}}
}

Downloads last month: 86

Total file size:

25.5 GB

Request access to EgoCrowd-500H (preview)