Visual-data procurement should specify images, source records, generation parameters, labels and quality checks separately. The number of images only records how many image files meet the stated counting rule. It is insufficient evidence of complete labels, training readiness or improved field performance.
List deliverables by their content
| Deliverable | Details to specify | What it does not independently establish |
|---|---|---|
| Captured images or video | Device, scene, format, time and original-data scope | Coverage of every field condition |
| Synthetic or augmented samples | Origin, generation method, parameters and batch | Equivalence to the real field distribution |
| Class information | Class definitions, field meanings and confirmation method | Object locations or pixel-level labels |
| Detection, segmentation or character annotations | Label type, rules, version and checking method | Successful model training or deployment |
| Quality-check records | Checks, sampling, exceptions and treatment | Target-model performance |
Check origin, permitted use and licences separately
Keep capture or generation sources and the permitted scope of use, processing and delivery. Permissions depend on documentation for each source. Possession of files is not permission for every subsequent use. Control public disclosure of customer scenes or identifiable information according to the agreed scope.
Accept representative examples first
Define the target task and label rules, then review representative examples. Check readable files, agreed formats and dimensions, matching labels, consistent class meanings, and traceable source and version metadata. Record unconfirmed labels or coverage as outstanding items.
Separate data acceptance from model evaluation
Accept data against completeness, quality, coverage and label consistency. Model performance additionally requires defined training, validation and independent test data, versions and metrics. Without those tests, delivery of synthetic images is insufficient evidence of completed training or improved performance.
Related services: data, training and evaluation and project enquiries.