What we do
Image and video
More than half of our work, and the oldest: the first Nairobi cohort drew boxes around weeds for a European agricultural client in 2016.
What we label
| Type | Used for | Typical accuracy target |
|---|---|---|
| Bounding boxes | Object detection: products, vehicles, people, animals | IoU ≥ 0.85 |
| Polygons and segmentation | Exact outlines: crops and weeds, packaging, road surface | IoU ≥ 0.90 |
| Keypoints and pose | Body pose, head direction, hands, livestock | ≤ 4 px mean error |
| Tracking | Objects and people followed across video frames | ID switches ≤ 1 per 1,000 frames |
| Attributes | Colour, state, occlusion, visibility, text on signs and labels | ≥ 98% exact |
| Classification | Whole-image and whole-clip labels | ≥ 98.5% agreement with gold |
- Targets are agreed per project and measured against gold sets; see Quality. IoU is the overlap between a labeller’s shape and the reference shape.
Video and first-person imagery
Video is labelled frame by frame or on keyframes with interpolation, depending on what the model needs and what the client will pay for. Labellers track objects and people across a clip, mark events with start and end times, and flag frames that are too blurred to label.
We also label first-person (“egocentric”) video and stills from wearable and head-mounted devices, including clinical devices, for activity recognition, gaze and attention research and assistive technology. Projects of this kind usually involve bystanders and private places, so they run only in our restricted rooms, with tier S2 or S3 wellbeing provision, and with redaction rules agreed before work starts.
Tools
Most projects run in Rungwell Tasks, our own labelling platform, which exports to the usual formats (COCO, Pascal VOC, YOLO, and JSON to your schema) and keeps a full audit trail of who labelled and reviewed each item. If your team has its own tool, we can work in it, provided it supports per-labeller audit and our labellers can reach it from our centres.
Throughput
A trained team of 40 labellers draws roughly 60,000 to 90,000 boxes a shift on uncluttered images, or 4,000 to 6,000 detailed polygons. We will give you a real estimate after the pilot, from your data, rather than a number from a brochure.
First-pass acceptance across all our projects was 98.7% in 2025.
Page last reviewed 28 September 2026