ANNOTATION & DATASET DESIGN
Turn hard examplesinto clear labels.
Specialist annotation for complex video and multimodal AI tasks.
Agree the rules before annotation.
Set label definitions, edge-case rules, reviewer agreement and export format before production.
What we label
01
Video and temporal
Actions, events, boundaries and sequences.
02
Spatial and visual
Objects, regions, relationships and interleaved visual-text outputs, where specified.
03
Human judgement
Specialist review, rubric design and difficult-case adjudication.
04
Quality
Calibration sets, sampled review, issue taxonomies and structured delivery.
What we deliver
- 01A working annotation guideline
- 02Representative accepted examples
- 03Structured labels in your schema
- 04A documented QA process
Inside an annotation record
From event to reviewable record.
This schema sample is synthetic. For production, we agree the ontology, time boundaries, entity links and edge-case rules with your team.
Source clip
clip_0042.mp4
EGOCENTRIC
00:12—00:16
Select annotation layers
3/3 visible
12.0s13.0s14.0s15.0s16.0s
actor
right_hand_01
contact
handle_01
event
open_drawer
Edge-case rule: contact begins at the first visible touch, not when the object starts moving.
Structured output
JSON / V1{ "event": "open_drawer", "interval_s": [13.12, 14.16], "actor": "right_hand_01", "target": "handle_01"}Fields shown
actor · contact · event
Start with a fixed-scope pilot.
After a no-charge scoping call, test the method on hard examples for 1–2 weeks. Keep the annotated sample, method memo and QA report whether you scale or stop.