# Annotation instructions v1.0

These are the instructions the sample labels were made to. A project brief replaces them with rules agreed with the lab.

## What to label

Label every object the task acts on and the effector that acts on it. In this clip: bottle, cup, right hand.

## Boxes

Draw the tightest axis-aligned box around the visible extent of the object. Do not extend a box over parts hidden behind another object. Boxes are `[x, y, width, height]` in pixels, origin top left.

## Keyframes

Place a keyframe every 0.5 s while an object is still or moving slowly, and every 0.25 s while it is moving or rotating. Positions between keyframes are linearly interpolated. Each object keeps one track id for the whole clip.

## Action segments

One segment per action, contiguous, with no gaps. A boundary goes on the first sampled frame where the motion changes: the bottle starts to rotate, stops over the cup, or starts back.

## Ambiguous cases

| Case | Rule |
| - | - |
| Occlusion | When the hand hides part of the bottle, box the visible part only. |
| Effector extent | The hand box ends at the wrist joint. The forearm is not labelled. |
| Idle effectors | An effector that takes no part in the task is not labelled. The left hand is in frame throughout and has no track. |
| Transparent objects | The cup is boxed on its rim and base outline, not on reflections on the table. |
| Unobservable events | Liquid is not visible at this resolution, so the pour segment is defined by bottle pose over the cup, not by flow. |
| Boundary tolerance | Segment boundaries are accurate to the sampling step, plus or minus 0.25 s. |

## Review method used on this sample

Two passes by one annotator. Pass one drew boxes at 0.5 s. Pass two burned the interpolated boxes onto 28 of the clip's 192 frames (every 0.5 s to 2.5 s, then every 0.25 s through the motion), compared each against the frame, and set a measured keyframe at every frame checked on the bottle and hand tracks. The cup did not move and keeps 0.5 s keyframes. The 8.0 s keyframe was checked against the clip's last frame, at 7.96 s. Any edge 4 px or more off was logged as a correction. Nine corrections were made; they are listed in `labels.json` under `review.corrections`.

This is a self-check, not an independent review. There is no second annotator, so no agreement score or IoU is reported. Boxes were measured by eye against a 60 px grid; expect errors of a few pixels.