1. Collect and label
Add authorised pictures and draw boxes around the target objects. Include realistic lighting, camera angles, partial visibility and background-only scenes. Keep provenance and usage rights with the data.
Sentinel ID · National identity & operationsExplore all four products →
AI training
Improve detection with representative examples
Add useful examples, mark the objects you want to detect and check the results before using a new model at a site.
2. Add to existing training
Add more examples to existing labels or continue from a selected earlier model. Training can reuse prior examples so that new data does not become the entire training set.
3. Evaluate independently
Hold validation and test pictures outside training. Review per-label precision and recall, missed objects, duplicate detections and false alarms on full frames.
4. Review before activation
Review the test results before activating a new model. It must meet the system’s validation requirements. A failed training run leaves the previous model in place.
Test each detector with the cameras and conditions where it will be used. Review false alarms and missed detections before approving it for live monitoring.
Training has a controlled purpose
- Detector-training permissions limit who can change datasets and models.
- Training runs locally using the installation's configured resources.
- Data quality, diversity and representative testing determine practical usefulness.
- Biometric candidates and object detections remain subject to human review.
Plan the acceptance test
Define the labels, camera positions and acceptable review workload. Measure false alarms and missed detections using scenes from the intended deployment. Extend the dataset when failures show a coverage gap.
Site security workflow Discuss training needs