Bricklaying video annotation guide

Task SCW-BRICK-01 · Version 1.0 · 2026-09-21

1. Task and delivery

ItemRequirement
Sourcebrick_work_01.mp4 (sent separately — this website does not host it). Head- or chest-mounted first-person view. Annotate the first 9 minutes as 18 clips of 30 seconds each.
SubjectThe camera wearer only. Other workers may be used as safety evidence, but do not count toward the wearer’s work time or brick count.
Stage 1Independently fill four ground-truth fields for all 18 clips. Use “Cannot tell” when the picture is not enough.
Stage 2After stage 1 is complete, score 59 anonymous activity summaries against the original video.
DeliveryClick Check and submit. The site stores your JSON for the study lead. Also download the file as a backup. Drafts are not a final delivery.
RolesAnnotators work independently. The study lead aggregates. Disputed actions or safety labels go to a business reviewer — do not guess.

2. How to work

  1. Get brick_work_01.mp4 from the study lead (download link or file transfer). Open this site in Chrome or Edge, then load that local file — or paste the video URL. The site only accepts the ~9-minute task video.
  2. Enter your name or ID and open this guide. Skim the video once, then fill clip by clip.
  3. Use Play this clip. Use 0.5× or ±1 s when you need to see a motion. Times in the form are 0–30 seconds inside the clip, not the full-video clock.
  4. Fill all four sections, tick the check box, then Check and finish this clip. Missing fields or non-contiguous state intervals will block you. Finished clips turn green.
  5. After all 18 clips, open Anonymous summary scoring. Rewatch the matching clip and score each summary with a short reason.
  6. Click Check and submit. Keep the downloaded JSON.

Saving: The page also stores a draft in this browser. Export a draft before you switch computers. Import that JSON to resume. Private/incognito windows and cleared cache lose the local draft. Do not share one browser draft across annotators.

3. The four annotation fields

1. Activity summary

Write the main activity and the actions that actually happen or finish, usually 1–2 sentences. Use ordinary language. Distinguish preparing, in progress, and completed.

What you see (examples, not answers for this video)How to write it
The mortar pan is lifted, but nothing is pouredWrite “carrying the mortar pan”; do not write “poured mortar”.
A brick already in place is tapped or shiftedWrite “adjusting a brick”; do not count a new placement.
Hands are occluded; the motion is unclearChoose “Cannot tell” and say what is blocked. Do not reconstruct the full sequence from context.

2. Work state

Split the clip wherever the state changes. Intervals must cover 0–30 s with no overlap and no gap. Mark unclear parts as “Cannot tell”.

StateUse when
WorkingVisible task behavior: laying bricks, spreading mortar, measuring, fetching, carrying. Getting a tool or moving with materials for the task counts as working.
IdleClear evidence of stopping, waiting, or resting. A still camera, hands off-screen, or another person working is not enough to call the wearer idle.
OtherA clearly visible non-task action.
Cannot tellYou cannot tell working / idle / other. Do not infer intent from a head turn or unexplained movement.

Evidence should be a visible action, not an unevidenced “waiting for materials”. Times may be rounded to about 1 s. Format example only: 0–10 working, 10–18 cannot tell, 18–30 working. That is not an answer for this video.

3. Brick count

One placement: the wearer puts a new brick into its target position and lets go. Use “Mark current time” or type the release time. The page counts the timestamps.

4. Safety observations

Every category needs Present / Absent / Cannot tell. If Present, add event type, times, severity, and evidence. One continuous event in a clip is one row. Events that cross clips are marked in each clip; the lead merges them later.

TypeMarkDo not infer
Trip hazardDebris, obstacles, or trip objects in a walkway, and where they are.Materials in frame are not automatically a path risk.
Work at heightVisible platform, edge, or drop.Height alone does not prove missing fall protection.
Hot workVisible welding, cutting, or similar.A tool in view is not hot work in progress.
Missing PPERequired PPE for this task, with the body part clearly visible and clearly missing.Head or harness off-screen → Cannot tell, not Absent.
Quality concernA quality issue you can point to in the picture, with a reason.If you do not know the spec, leave it for review. Do not fail quality on impression.
SeverityMeaning
Scene noteYou can confirm the setting, but not a risk. Example: a high platform, recorded as scene only.
Needs reviewThere is picture evidence, but danger or protection still needs a reviewer.
Clear visible riskThe picture itself shows the risk, and you can write what is visible.

These are experiment labels, not legal, site-compliance, or acceptance decisions. If unsure, use Cannot tell or Needs review and say why.

4. Scoring anonymous summaries

ScoreStandard
PassMain activity is right; key completed actions are not invented or omitted. Near-synonyms are fine.
FailWrong actor or main action; a prep step is stated as finished; a key completion is missing; or extra detail has no evidence. Cite the sentence and video time.
Video insufficientA human also cannot confirm from the picture. Describe the occlusion. If the summary asserts a completed action with no evidence, score Fail — do not use this option to excuse it.

Write a short reason even for Pass. Do not pass a summary because it sounds fluent or because you guessed the model. If ground truth must change, go back to stage 1 and tell the lead. Do not edit truth to match a model.

5. Delivery checklist

Review: If two people annotate, they submit separately and do not share answers. The lead sends disagreements to a business reviewer and keeps both the originals and the ruling. If there is no second annotator, the report should say single-annotator.

Footage: Verbose Tech Labs LLP, CC BY 4.0. Source: skilled-commercial-work-egocentric-sample. This tool does not call a model. A compact copy of the clip is streamed for annotation; it is not a substitute for the original 1080p file in archival work.