Computer Vision Data & Performance Partner
Reliable data. Trained models. Measured improvement.
We build high-quality training datasets and run annotation at scale for in-house vision engineers, and we train complete models for companies that have no computer vision team of their own.
- Annotation & workflows
- Ambiguous cases
- Model training
- Performance experiments
Detection model. Finds and localises every object, each with a confidence score. Low-confidence predictions are routed back for review and relabelling.
Ordering model. Ranks objects along a scale, for example fruit by rot severity, instead of forcing a yes/no class.
Classification model. Assigns one class per object, for example accept or reject on a sorting line.
Clients
Computer vision teams we work with
Our specialisations
Deep expertise where data and model performance meet.
Engage Q2Label for one focused work package, or connect all four disciplines into a single systematic improvement loop.
What we doWhat you receive
01 / PRECISION DATA ANNOTATIONS & WORKFLOWS
Reliable training data at scale, without the operational overhead.
We design and operate efficient image and video annotation workflows: label ontologies and annotation rules, dataset composition, task instructions, workforce management, model-assisted pre-labelling where appropriate, and documented quality review.
What we do
- Image & video
- Boxes, polygons, masks, keypoints
- Label ontology
- Annotation guidelines
- Dataset composition
- Model-assisted pre-labelling
- Reviewer calibration
- Batch review & rework
What you receive
- Annotated dataset
- Task specification
- Quality-review results
- Agreed export format
02 / AMBIGUOUS CASES HANDLING
Decide the hard cases once, then label them the same way every time.
Borderline, unclear and partly visible objects are where labels start to diverge. We write explicit rules for them, detect disagreement between annotators and route it to expert judgment. For ordered properties such as clarity, severity or growth, we structure richer ordinal labels and, when justified, expert-relative comparisons.
What we do
- Ambiguity rules
- Visibility & occlusion
- Multi-annotator voting
- Expert escalation
- Ordinal scales
- Calibration anchors
- Expert-relative comparison
What you receive
- Edge-case rubric
- Disagreement overview
- Graded labels
- Validated pilot workflow
03 / MODEL TRAINING
From reviewed dataset to a trained model your team can run.
We train and fine-tune computer vision models on your data, for detection, segmentation, classification and grading tasks. Runs are tracked and reproducible, and the weights, configuration and evaluation are handed over to your engineers.
What we do
- Detection
- Segmentation
- Classification
- Ordinal grading
- Fine-tuning
- Train / val / test splits
- Augmentation
- Tracked runs
- Export for deployment
What you receive
- Model weights
- Training configuration
- Run log
- Hold-out evaluation
04 / MODEL PERFORMANCE ENGINEERING
Find the bottleneck. Test the intervention. Measure the result.
For models we trained or models you already run: baseline evaluation, failure analysis, controlled data and training experiments, and reproducible comparisons on independent data.
What we do
- Baseline & metrics
- Per-class error analysis
- Scenario slicing
- Label defect checks
- Controlled experiments
- Version tracking
- Independent hold-out
What you receive
- Evaluation report
- Error taxonomy
- Experiment log
- Recommended next steps
LABEL → DEFINE → TRAIN → MEASURE → REPEAT WHERE IT PAYS OFF
Technical examples
What careful computer vision data work looks like.
Two simplified examples that make ordinal grading and quality control tangible.
- Half ripe
- Unripe
- Fully ripe
- Turning
- Ripe
- Severe
- Light
- Very severe
- Healthy
- Moderate
- Minor
- Rotten
- Moderate
- Sound
- Severe
Ordinal grading by ranking
Ripeness, leaf disease and rot have no sharp class boundaries. Instead of forcing a yes/no label, samples are ordered relative to each other, from less to more ripe or severe. The ranking gives a consistent five-step scale, and the threshold for harvesting, treating or rejecting can be set afterwards without relabelling.
DEFECT CANDIDATE · SCHEMATIC
- AI pre-labelPimple defect
- Annotator 1Pimple defect
- Annotator 2Surface stain
- Disagreement detected → escalate to expert
- Expert reviewPimple defect ✓ final label
Quality control with multi-annotator voting
A QC workflow from food processing: start from an AI pre-label, compare it with two independent annotators, and trigger expert judgment when they disagree. Annotation stays scalable while label quality on ambiguous samples is protected.
How we work
Engineering discipline, clear ownership and evidence at each step.
We adapt to the client’s existing computer vision stack. A pilot validates the approach before work is scaled or experiments are repeated.
- 01 / ALIGN
Define the goal
What must the model detect or assess, under which constraints, and how will success be measured?
Baseline & success metric - 02 / SPECIFY
Design the workflow
Label rules, datasets, quality controls and the experiment protocol, checked with your experts.
Annotation specification - 03 / PILOT
Run a controlled pilot
A representative batch or experiment. Edge cases are inspected before any larger commitment.
Reviewed pilot batch - 04 / EVIDENCE
Scale and measure
Reviewed data, trained models or experiment results, with open risks documented.
Results & next decision
Deliverables, not just recommendations.
Scope-specific technical outputs that your engineers can review, reproduce and use in their own pipeline.
- Annotation specifications
- Reviewed training datasets
- Quality-control summaries
- Label ontology / decision rules
- Trained models & training configs
- Evaluation breakdowns
- Experiment summaries
Why Q2Label
Specialist support for teams that already know computer vision.
We contribute where another perspective, a dedicated data operation or a controlled experiment can unlock the next engineering decision.
Cross-project perspective
Repeated exposure to different datasets, annotation ambiguities and model behaviours informs how we approach each new problem.
Data work connected to model outcomes
Annotations, definitions and sampling decisions are guided by what the model must learn and how its performance will be evaluated.
Work alongside your engineers
Domain knowledge stays with your experts. We bring dedicated specialist capacity, documented methods and a clear division of responsibilities.
About Q2Label
A focused engineering partner, based in the Netherlands.
Q2Label was founded in January 2024 in Eindhoven. We combine computer vision expertise with managed annotation operations, and work for computer vision teams in the food and agri sectors.
We focus on the practical interface between datasets, ground-truth quality, model training and evaluation. Our team is intentionally specialised: clients work directly with the people who design the workflow and review the technical decisions.
Bart de Backer
Co-founder · Workflow expert
Works on annotation strategy, project delivery, operational coordination and quality-control workflows.
“We’ve seen how unclear instructions ruin datasets. That’s why we lock quality before scaling.”LinkedIn
Erik van der Poel
Co-founder · Computer vision engineer
Works on model training, evaluation, experiment tooling and research into expert-aligned visual grading.
“We understand the importance of qualitative data to improve your model performance at experimentation and in production.”LinkedIn
Let’s talk
Bring us a real computer vision challenge.
Tell us which part of your dataset or model workflow needs attention. We can define a focused first assessment together with your engineers.
We use these details to respond to your enquiry. Do not include confidential images or personal data in this form.