Dataset Generation

Dataset generation supports active learning loops for improving detector models.

What It Does

• Scores frames using configurable quality metrics.
• Selects challenging frames for annotation.
• Exports image/label artifacts for downstream training.

Quality Metrics (Conceptual)

• Low confidence detections
• Detection count mismatch vs expected target count
• High assignment costs
• Track loss events
• Optional uncertainty-based triggers

When to Use

• Detector underperforms on specific lighting or behaviors.
• You need focused retraining data instead of random frame sampling.

Tradeoffs

• Aggressive selection can bias toward outliers.
• Conservative selection may miss edge cases.

Practical Loop

1. Run MAT and generate candidate frames.
2. Validate/annotate selected frames.
3. Retrain YOLO model.
4. Re-run MAT on representative videos.
5. Compare confidence and identity continuity metrics.