Computer Vision Data And Labeling

Computer Vision Data And Labeling

Build a visual dataset whose sampling unit, labels, lineage, and split structure match the deployed task. Treat annotation guidance and collection policy as versioned model inputs, not informal preparation work.

Core Rules

• Define the production population and correlation unit before collecting or splitting: subject, object, scene, event, camera, source, template, site, or time window.
• Version the label ontology and guidance with edge cases, examples, adjudication, provenance, and the dataset versions they apply to.
• Audit class, source, camera, scale, position, quality, and missing-label coverage; an aggregate sample count is not coverage evidence.
• Transform every spatial or temporal label consistently with the input. Inspect the resulting image-box-mask-keypoint-track pairs, not only augmentation code.
• Keep manual, imported, auto-generated, and corrected annotations distinguishable. Confidence from an auto-labeler is not ground truth.

Boundaries And Handoffs

This skill owns visual acquisition, dataset and annotation contracts, split integrity, data quality, and the data-improvement loop. Use $computer-vision-modeling-and-training for architecture and training decisions, and $computer-vision-evaluation for model metrics, thresholds, robustness evidence, and release gates. Use $ml-system-design for generic pipeline orchestration and lifecycle concerns.

Rights, consent, retention, access, deletion, tenant isolation, and audit policy are system constraints; use $system-design for their technical enforcement and do not invent legal conclusions. Security testing must remain authorized and defensive.

Reference Routing

• Read collection, lineage, and coverage for acquisition plans, metadata, sampling, and dataset acceptance.
• Read annotation contract and quality for labels, guidance, calibration, agreement, adjudication, and audits.
• Read splits, leakage, and augmentation for grouping, near-duplicates, temporal data, transforms, and immutable tests.
• Read iteration and feedback for imbalance, active learning, auto-labeling, synthetic data, and production feedback.

Workflow

1. Define the target population, sampling unit, label unit, metadata, and exclusions.
2. Design the annotation contract and run a small calibration and adjudication pass.
3. Inspect coverage and correlations, then create grouped train/validation/test splits.
4. Establish a simple dataset and model baseline before scaling annotation automation.
5. Add augmentation, active learning, auto-labeling, or synthetic data only with provenance, fixed evaluation slices, and human or golden-set audits.
6. Version the resulting dataset manifest and record owners, known gaps, and the next collection decision.

Output

Include population and capture plan; sampling and label units; ontology and guidance; QA and adjudication; metadata and provenance; split and leakage controls; coverage and imbalance analysis; transform policy; automation/synthetic-data controls; versioning, rights/privacy owner, feedback loop, risks, and acceptance checks.

Quality Bar

• Do not randomly split correlated frames, scenes, subjects, templates, or duplicates.
• Do not scale annotation before calibrating the specification on ambiguous cases.
• Do not use confidence-only sampling without checking calibration and diversity.
• Do not let a changing auto-labeler silently redefine the dataset.