Delivery specification
How your dataset is delivered
Files, naming, index and handoff, so ingestion can be wired up before the first clip ships.
- Container
- MP4 or MOV
- One file per recording
- Video
- Stereo, side by side
- 2×1920×1200 native, 30 fps
- Delivery unit
- One recording, one file
- One task, or one session
- Naming
- Structured
- Country, setting, task, IDs
- Index
- CSV or JSON
- One per delivery batch
- Handoff
- Pull or push
- Your cloud or ours
01
What a single delivery contains
A standard delivery is raw data: video files, per-clip metadata, one index and the schema. Each video is one complete task; hands, tools and work surface stay in frame.
- Video files. One MP4 or MOV per task, camera-native and untouched. Clip length to your spec.
- Per-clip metadata. Capture context and technical attributes.
- One index. CSV or JSON: every video with its task, setting, duration, size and consent state.
- Schema docs. The index fields and task taxonomy.
Every recording passes Capture QC before it ships (section 08). Labeling and Dataset QA are optional (sections 06 and 08).
02
One continuous recording, one file
One continuous recording is one file and one unit of QC; almost always a single task.
- No mid-task cuts. One uninterrupted take, never spliced or trimmed in the middle.
- Segments are ours, not yours. Long recordings travel in segments and are joined before delivery.
- Usually one task; sometimes a session. Related activities in one take stay in one file, with per-task time ranges in the index.
- Separate takes, separate files. Told apart by capture number.
Built-in variety
One contributor can submit about 5 hours of the same activity in the same setting; past that they move to another task or environment.
03
The filename tells you what's inside
Filenames are generated from capture data, so they are consistent and safe to parse. Six fields:
CC_Region_EnvKind_Task_NNN_NNNN.mp4
VN_HaNoi_industrial_assembly_007_0042.mp4
Contributor 007's 42nd capture: assembly, industrial setting, Ha Noi, Vietnam.
- CC
- ISO 3166-1 alpha-2 country code (VN, ID).
- Region
- Province or city, ASCII only (Ha Noi → HaNoi).
- EnvKind
- home, commercial, outdoor or industrial.
- Task
- Task category; the sub-task is in the index.
- NNN
- Pseudonymous 3-digit contributor number.
- NNNN
- 4-digit capture number, unique per contributor.
04
Folder layout
Each video sits in a folder keyed by an opaque capture ID; index and schema at the root. Names carry nothing personal, and every file has a stable address.
<delivery-root>/
dataset/
videos/
a3f9c1…/ # one folder per capture (opaque ID)
VN_HaNoi_industrial_assembly_007_0042.mp4
b7e204…/
VN_HaNoi_industrial_inspection_007_0043.mp4
index.csv # or manifest-<batch-id>.json
schema.md # index fields and task taxonomy
addons/ # optional: labels, annotations (section 06)
qa/ # optional: Dataset QA report (section 08) 05
Your ingestion index
One index per batch, CSV or JSON, with structured facts and pseudonymous numbers only.
{
"export_id": "b3d1…",
"created_at": "2026-10-18T09:12:04Z",
"items": [
{
"filename": "VN_HaNoi_industrial_assembly_007_0042.mp4",
"uri": "<delivery-root>/dataset/videos/a3f9c1…/VN_HaNoi_…_0042.mp4",
"duration_seconds": 512.4,
"size_bytes": 843127740,
"task_category": "assembly",
"task_subcategory": "fastening_housing_screws",
"environment_kind": "industrial",
"province": "Ha Noi",
"recorded_at": "2026-10-16T02:31:10Z",
"stereo_layout": "side_by_side",
"resolution_per_eye": "1920x1200",
"fps": 30,
"contributor_number": 7,
"capture_number": 42,
"consent": { "recorder": true, "third_party": true, "workplace": true }
}
]
} - filename
- File name (section 03).
- uri
- Path within the delivery.
- duration_seconds
- Runtime after trims.
- size_bytes
- File size on disk.
- task_category
- Task type; matches the filename.
- task_subcategory
- Specific activity.
- environment_kind
- home, commercial, outdoor or industrial.
- province
- Region.
- recorded_at
- Capture time (UTC).
- stereo_layout
- Eye packing in the frame.
- resolution_per_eye
- Native pixels per eye.
- fps
- Frame rate.
- contributor_number
- Pseudonymous ID.
- capture_number
- Per-contributor sequence number.
- consent
- Recorder, third-party, workplace flags (section 09).
Stable across batches; we can add fields or emit your schema.
06
Optional add-ons
Raw video, per-clip metadata, the index and schema docs ship with every delivery. Labeling and annotation are optional, before or after delivery depending on the project. On request, in addons/:
- Task labels to your taxonomy
- Frame-level annotations
- Bounding boxes and keypoints
- Temporal segmentation
- Custom file structures and naming conventions
- Train / validation / test splits
addons/
labels/
annotations/
keypoints/
segmentation/
splits/ 07
How you receive it
Two handoff models on signed, access-controlled storage. Same files either way.
Pull · default
Read from our bucket
A dedicated bucket, read-only for your service account. Pull on your schedule.
Setup: one service-account identity from you.
Push · optional
Delivered to your storage
Each batch written into your own S3 or Azure Blob bucket.
Setup: write credentials from you.
Multi-part files or a specific index schema? We align before the first batch.
08
Only checked footage ships
QC checks the raw footage at capture and is part of every delivery. QA verifies the finished dataset and is optional; many buyers run their own.
- Capture QC · always. A reviewer watches the whole task against a fixed rejection list and may trim dead time at either end. Failures are excluded.
- Diversity control · always. The ceiling from section 02 is enforced at QC.
- Dataset QA · optional. Automated checks flag; a second reviewer resolves each flag. A confirmed problem rejects the whole clip.
When Dataset QA is ordered, its report ships in qa/.
qa/
spot_checks.csv
label_audit.json
completeness_report.md 09
Privacy and consent, by construction
No personal data travels with the dataset; consent is recorded per file.
- No PII in filenames or the index. Contributors are numbers only.
- Consent recorded per capture. Three flags: recorder, third party, workplace.
- Consent-gated pipeline. Captures without consent never enter a batch.
10
Formats
| Video container | MP4 or MOV, as written by the camera |
|---|---|
| Video codec | Camera-native file, untouched (raw-first). H.264 MP4 on request. |
| Resolution | 2×1920×1200 native (per eye) |
| Frame rate | 30 fps |
| Stereo layout | Side by side, left | right |
| Frame sync | Both eyes ship in one side-by-side frame, so no separate sync file is needed. Raw per-eye streams with the camera's frame-sync code are available on request. |
| Other devices | 1080p to 4K, 30–60 fps, MP4 (H.264) on standard devices |
| Metadata format | JSON, per clip |
| Index format | CSV or JSON, per batch |
| Capture rig | Head-mounted stereo camera by default. Smartphone, multi-cam, wrist-cam and exo/egocentric configurations on request. |
11
The stereo camera
The default rig: a head-mounted stereo camera with an integrated IMU, both eyes in one frame-synchronous stream.
| Configuration | Stereo (dual sensor) with integrated IMU |
|---|---|
| Sensor resolution | 1920 × 1200 per sensor |
| Stream | 4000 × 1200 at 30 fps (combined stereo UVC stream) |
| Composition | Two 1920 × 1200 sensors side by side (3840 × 1200) plus a 160 × 1200 IMU/barcode data strip |
| Pixel size | 3.0 µm × 3.0 µm |
| Field of view | 160° diagonal (H 134°, V 72°); configurable from about 100° to 200° |
| Stereo baseline | 60 mm |
| Output format | MJPEG / YUV2 (YUYV); H.264 via onboard encoder |
| Focus | Fixed |
| Shutter | Global shutter |
| Centre clarity | 800 LW/PH |
| Signal-to-noise ratio | 37 dB |
| Dynamic range | 70 dB |
| Sensitivity | 1.4 V/lux·s at 550 nm |