Request access to the Outpost Labs egocentric evaluation sample
Outpost Labs reviews requests individually and confirms the evaluation licence and authorised Hugging Face account before granting access.
The sample contains original, unblurred footage. Use is limited to internal technical evaluation under LICENSE-EVALUATION.md.
Log in or Sign Up to review the conditions and access this dataset content.
outpost-ego-v0-eval
Evaluation only. This pack is licensed by Quinn Risk Management, Inc. d/b/a Outpost Labs, for evaluation only, under
LICENSE-EVALUATION.md. Your rights are exactly those in that licence; nothing in this card
grants or extends them.
Summary
149 video clips (654.2 minutes) of real work recorded on site in India and the United States. The five site codes are in-hyd01, in-kadapa01, in-pdy01, in-vjw01, and us-nyhome01. The lanes are garment-ironing, home, machine-shop, pharma, phone-repair, tailoring, tile-setting, and woodworking. Each clip ships with audio removed, IMU samples on the video clock, and per-clip IMU-to-video sync values.
Of the 149 clips, 14 show electronics repair from a camera fixed above the workbench
(the phone-repair lane, with capture.mount set to fixed-overhead). These clips
show the type of work. The remaining 135 clips were filmed with head-worn cameras.
If your use case requires a head-worn view for electronics repair, we'll have
operators wear the camera and can collect that footage at scale.
| Lane | Clips | Minutes | Delivered fps | Sites |
|---|---|---|---|---|
| garment-ironing | 17 | 85.0 | 60 | in-kadapa01 |
| home | 25 | 118.9 | 30, 60 | in-hyd01, in-kadapa01, us-nyhome01 |
| machine-shop | 27 | 119.3 | 60 | in-hyd01 |
| pharma | 23 | 81.0 | 30, 60 | in-vjw01 |
| phone-repair | 14 | 53.4 | 30 | in-hyd01 |
| tailoring | 8 | 37.4 | 60 | in-kadapa01 |
| tile-setting | 24 | 119.1 | 30 | in-hyd01 |
| woodworking | 11 | 40.0 | 30 | in-pdy01 |
Layout
README.md this card
LICENSE-EVALUATION.md the evaluation licence
manifest.csv one row per clip: clip_id, sha256, bytes, lane, site, worker,
consent_form_version, deid_version, sync_residual_ms, calibration_id,
licence_tier, fps, frames, duration_s, source_fps, keep_every,
registration, source_file, trim_start_s, trim_end_s
SHA256SUMS SHA-256 of every file in the pack
docs/ known_limitations, rights_provenance, calibration, label_spec, qa_summary
catalog/ episodes.jsonl, session_episodes.jsonl, segments.jsonl if produced
raw/<site>/<worker>/<session>/clip-<id>.mp4 original video (not blurred), 60 fps or 30 fps as recorded, no audio
clip-<id>.json catalog fields, the shipped range, sync
clip-<id>.imu.parquet IMU samples on the video timebase
clip-<id>.annotations.jsonl step segments, where produced
clip-<id> uses the first 12 hex characters of the BLAKE3 hash of the clip's MP4 bytes.
The worker field and path component use the placeholder not-provided. Session IDs
use <camera>.sNNNN, where NNNN is the session's first file counter or the next
available number if that counter is already in use.
Video, IMU and sync
- Video. Each clip is the camera's own HEVC (H.265) video stream, copied into a new MP4 without re-encoding: every video packet is byte-identical to the camera's original, and no frame is decoded and re-encoded, blended, interpolated, re-timed, stabilized or rectified. Clips keep their recorded rate and resolution (69 clips at 30 fps, 80 clips at 60 fps); every recorded frame of a clip's shipped range is in its MP4, at its original presentation time (the first shipped frame is 0 s). The camera's container metadata (its clock and vendor names) is not copied, but the original file name (
source_file) andcapture_datestill carry the camera clock (Known limitations). Each clip's JSON recordsvideo.fps,video.framesandvideo.integrity. - IMU. Gyroscope (rad/s) and accelerometer (g) samples at about 799 Hz on the video timebase
(
t_video_s, using the clip's fitted offset and drift), within_videomarking samples inside the video's span. - Sync. A per-clip IMU-to-video offset and clock drift, fitted by cross-correlation of gyro and visual motion (
xcorr_inhouse), with its residual (sync_residual_msinmanifest.csv, details in each clip's JSON). 148 of 149 fits have the top quality grade (per-clip grades indocs/qa_summary.md).calibrated: falsefor every clip: the sync was fitted without a lens calibration.
Session episodes
- What a session episode is. The camera writes a new file every 5 minutes, so one task often
runs across several clips.
catalog/session_episodes.jsonlholds one record per session episode: the shipped clips of one recording session that run back to back, in order. It is a grouping, not a file: no video is joined or re-encoded and each clip stays its own file. Each member gives its shipped range of its recording in seconds (source_start_s,source_end_s) and in delivered frames (frame_range, half-open), and its place in the episode (episode_start_s,episode_frame_range: the members' shipped lengths and frames laid end to end). Each clip's JSON names its episode (session_episode).session_episode_idis<session_id>.c<NNNN>, from the first member's file counter, so a clip withdrawn from a later version does not renumber the other episodes. It is not theepisode_idofcatalog/episodes.jsonl, which holds one record per clip and per task unit (docs/label_spec.md). A recording file picked in several pieces ships each piece as its own clip; an episode that starts at such a piece is<session_id>.c<NNNN>.k<N>, N the piece's place in its file, and two pieces of one file never join: the time between them was not picked. - How clips are grouped. Two clips join only when they are consecutive files (counters n and n+1) of the same session, lane and frame rate; the earlier file is a full split, at least 299 s long (the camera starts a new file only when one reaches 300 s, so a shorter file was stopped), and ships to its end, and the later ships from its start (each within 0.5 s); and the camera clock in the file names puts the later file's start within 1.5 s of the earlier one's end. Anything else starts a new episode, so every clip is in exactly one. Only shipped clips are grouped: a clip left out of the pack splits a run.
- What a join gives.
join.clock_gap_sis the camera clock's estimate of the time between the two files, in whole seconds (clock_resolution_sis 1), not a measured loss.join.trimmed_sis exact: the seconds the keep ranges leave out at the join. On bench evidence (measured on recordings at sixty frames a second) consecutive 5-minute files share one boundary frame (a split file holds 18,001 recorded frames in 300.0167 s): wherejoin.repeated_frameis true, the later clip's first frame repeats the earlier clip's last, so episode time laid end to end runs one frame long for each such join. To splice the clips on one clock, drop the first frame of each later clip whose join hasrepeated_frame. The IMU is per clip, with a gap of about 0.6 to 1.0 s at each join. - Episode ends.
start_boundaryandend_boundarysay why an episode starts or ends where it does:keep_mark(a keep range starts or ends there, more than 0.5 s from its file's start or end),file_edge(the neighbouring file of the recording exists but is not in this pack, so the episode is cut short),recording_startorrecording_stop(the recording itself began or stopped there), orunknown. A start-sidefile_edgewhose previous file has no measured length is inferred from the camera clock: the file names put that file at least a full split earlier, which a stop and a quick restart could also give. An episode is not a task: the rule checks recording continuity and lane, not what was done. Unlike a LeRobot episode, it has no single clock across its clips (each clip keeps its own video and IMU timeline) and no task description of its own (only its lane'stask). - In this pack. The 149 clips form 79 episodes: 27 join 97 clips (460.1 minutes, 7.7 h) and 52 are single clips. Episode ends: 45 keep_mark, 80 file_edge, 22 recording_start, 11 recording_stop.
Labels
The pack ships 4,925 step segments across all 149 clips in catalog/segments.jsonl
and the per-clip .annotations.jsonl files. Gemini (gemini-3.6-flash) drafted the
segments from original frames sampled at 1 fps. Segment boundaries are approximate
to about one second. A reviewer skimmed the drafts. Individual segments did not
receive human verification. Each shipped segment carries review_status: skimmed
and label_provenance: model.
Consent and rights
The licensor is Quinn Risk Management, Inc. d/b/a Outpost Labs.
Known limitations
This is a small evaluation sample. The footage is not de-identified. Sync was fitted without lens calibration, and step labels are machine drafts that received a human skim. See known limitations.
Out-of-scope use
- Any use that
LICENSE-EVALUATION.mddoes not expressly permit. - Any use beyond evaluation, and any further distribution:
LICENSE-EVALUATION.mdpermits neither. If a later written agreement permits either, the footage is de-identified first. - Any attempt to identify, re-identify or contact a person who appears in the footage.
Pack integrity
Check every file with sha256sum -c SHA256SUMS (or shasum -a 256 -c SHA256SUMS on macOS).
Production data service
Outpost Labs collects first-person video of skilled work on demand: new lanes and task lists, more wearers and sites, de-identified or original delivery, per-camera calibration, measured IMU sync and task-matched labels, each clip with consent and provenance records like this pack's. Terms for production data are agreed separately. Ask your Outpost Labs contact about a pilot (outpostdata.ai).
Special data: stereo lab capture (available on request)
- What it is. A separate section of 5 clips (4.9 minutes) of another kind of data that can be
collected in volume on request: a head-mounted stereo camera (left and right views) with an IMU on
the same hardware clock, recorded in a wet-chemistry lab in China (site
cn-szx-lab01). These clips are not counted in the clips, minutes, lanes, episodes ormanifest.csvabove, and are not registered in the Outpost catalog (no worker, session or device id). Files:special/stereo-lab/(its ownmanifest.csv); details:docs/special_stereo_lab.md. - Files. Each clip has
<clip_id>.left.mp4and<clip_id>.right.mp4(each camera's own HEVC stream, copied without re-encoding, audio removed),<clip_id>.imu.parquet(the main clips' IMU schema),<clip_id>.exposures.parquet(each recorded frame's exposure time and its index and container time in each MP4) and<clip_id>.json. - Timing. The camera timestamps every frame's exposure and every IMU sample on one clock (sync
method
hardware, no fitted offset): a frame every 16.640 ms (recorded at 60.1 fps), the IMU at about 601 Hz. 0 s is the exposure start of the earliest shipped frame of either view. The MP4 container times are the camera muxer's and jitter by a few milliseconds: pair frames and IMU byexposures.parquet, not by container time. - The two views start apart. A view can only start at one of its own keyframes without
re-encoding, and the left and right streams' keyframes do not coincide; in 5 of the 5 clips the
right stream also begins with frames that reference pictures not in the recording and cannot be
decoded, which are not shipped. Each clip's JSON gives each view's recorded frame range
(
views.*.source_frames). - Not labelled. No step segments or task units were drafted for these clips; each JSON's
labels.segmentsis null. - Not blurred; rights. Faces are not blurred. Quinn Risk Management, Inc. d/b/a Outpost Labs attested on 2026-11-13, in these words: "The capture partner confirmed that every person recorded in these clips consented to the recording and to its licensing to Quinn Risk Management, Inc. d/b/a Outpost Labs for evaluation."
- Downloads last month
- 25