Phased Strategy

Use a staged strategy to reduce risk and control compute costs.

Phase A: Baseline (no architecture change)

  • Evaluate current zero-shot prompt workflow (text=\"trees\") on labeled validation data.
  • Record baseline metrics: mIoU, Dice/F1, AP@50.

Phase B: Parameter-efficient transfer learning

  • Freeze most of SAM3.
  • Train either mask decoder only or LoRA adapters plus decoder.

Phase C: Targeted architecture changes

  • Add domain-specific modules.
  • Train new modules first, then consider wider unfreezing.

Phase D: Full or near-full fine-tuning

  • Unfreeze larger parts of the image encoder.
  • Use only if Phase B/C plateau and data plus compute are sufficient.

Phase E: Domain-adaptive pretraining (optional)

  • Continue self-supervised pretraining on large unlabeled local imagery.
  • Then run Phase B/C fine-tuning on labeled data.

See also:

PreviousTransfer Learning Fundamentals NextTransfer Learning Methods