Abelling
Managed data labelling and training-data preparation for AI teams. We turn raw images, video, LiDAR point clouds, text and audio into reviewed, training-ready datasets, so your engineers can spend their time on models rather than labelling queues.
Proposition
Reliably outsource your data labelling and training-data preparation to Abelling, so your team can focus on building models.
We are a managed service, not a self-serve tool. We agree the task and taxonomy with you, staff and train the annotators, review the work as it is produced and hand back data that meets the acceptance criteria you set.
- A team trained on your guidelines, including the edge cases that matter to your model.
- Review built into production, not added at the end.
- Delivery in the format and tool agreed for your project, from your own platform to MCAP data as used with Foxglove.
Supported tasks and sample outputs
Each service page sets out the tasks we offer, the review criteria we apply and the formats we deliver. The examples below come from programmes already described on this site. Any sample output we show is either permission-cleared or clearly synthetic, so you can see the labels, the rubric and the format side by side.
-
Annotated image, video, 3D and LiDAR data, from first-pass labelling to QA and correction of existing datasets.
- Boxes, polygons, segmentation, lines, keypoints and pose — used on programmes such as PPE compliance, workplace pose and workplace objects.
- Video and fisheye CCTV work, including queue head detection, retail surveillance, and published mass-transit masks and belongings.
- Instance segmentation for waste-stream classification, plus published identification work such as Japanese telephone poles.
- 3D and point-cloud annotation, from the published neuron and axon project to LiDAR QA and correction on MCAP data as used with Foxglove.
-
Annotated text and audio for language and speech models: classification, transcription, utterance collection, extraction and sentiment.
- Sentiment annotation, as in the published COVID sentiment story from tweets and Facebook posts.
- Audio classification, as in the published herding-dog audio project — tagging short recordings such as barking, howling, growling or silence.
- Transcription, utterance collection and entity extraction, scoped against written guidelines and reviewed before delivery.
-
Data operations beyond annotation, run by a managed team with review built in, either alongside labelling or as a standalone piece of work.
- Data entry, cleansing and enrichment: capturing, correcting and standardising records.
- Auditing, QA and model validation: checking existing datasets or model outputs against your rules, including correction of labels produced elsewhere.
- Data engineering and ETL, so labelled or cleaned data is delivered in a form ready for training or analysis.
-
The tools, formats and quality method we use to hand work back, including LiDAR workflows.
- Delivery in the tool and format agreed for your project, from your own platform to standard export formats.
- LiDAR and multi-sensor handling, including MCAP data as used with Foxglove.
- A documented review method, so you can see how quality is checked before acceptance.
How we work
Every project follows the same four stages, agreed with you before production starts:
- Scope and taxonomy: the classes, attributes and edge cases, written into guidelines.
- Calibration: annotators are checked against your edge cases before volume increases.
- Production and review: trained annotators work through the volume with review built in, often in two tiers.
- Acceptance and export: the work is checked against your acceptance criteria and exported in the agreed format.
LiDAR and multi-sensor work is supported, including point-cloud QA and correction on MCAP data as used with Foxglove. Read the full method in How We Work, or see tools and formats in Tools and Delivery.
Industries
Start from your sector. Our industry pages describe what we can do for each kind of team, including the data you are likely to hold, the labelling problems it raises and how we would approach them, not only the work we have already completed.
- Industrial and workplace safety: PPE, pose, near-miss and vehicle and pedestrian labelling from site CCTV.
- Autonomous vehicles and robotics: LiDAR QA and correction, camera and LiDAR fusion, and robotics episodes.
- Healthcare and medical annotation: clinician-led annotation for imaging, pathology and clinical text.
- Retail: spatial boxes and timestamped events on in-store CCTV.
- Recycling and environment: instance segmentation of waste streams at production volume.
- Warehousing: perception data for warehouse operations and robotics.
Selected stories
Three examples of recent work, each told in the same order: problem, data and task, tools, QA and delivery scale. Client names stay confidential.
- PPE compliance: more than 3.4 million region-of-interest tags for workplace-safety AI on industrial video.
- Queue head detection: QA and correction of more than 50,000 head boxes a month on overhead fisheye CCTV.
- Waste stream classification: about two years of monthly instance-segmentation batches for recycling AI.
Company and security
Abelling is an established data-labelling organisation, operating since 2018, with teams anchored in Dhaka, Chicago and Leeds, UK, and clients well beyond those locations. Read about Abelling.
Before a project starts, we agree with you in writing how your data is accessed, transferred, retained and deleted, who owns the outputs and how incidents are reported. See Security and Compliance.
Request a Project Quote
Tell us about your data, your labels and your quality target, and we will scope the work with you. A short brief covering modality, volume, taxonomy and timeframe is enough to start the conversation.