1. How is self-training implemented?
The platform has AI training capability. Users can upload defect samples collected by flight and train a detection model tailored to their own defect characteristics, enabling continuous model iteration. Training efficiency is positively correlated with compute cluster size; stronger compute means faster iteration.
2. How much compute investment is needed?
Forming baseline productivity requires roughly 1-2 million RMB in GPU investment. Two GPUs can run AI but slowly; adding GPUs multiplies training efficiency. Under cloud deployment, cloud compute can be used to reduce one-time hardware investment.
3. Self-training vs using the built-in model
| Method | Advantage | Investment |
|---|---|---|
| Built-in model | Ready to use, mature and stable | No extra investment |
| Self-trained model | Tailored to your defect characteristics | Requires compute + samples |
4. Applicable conditions and scenarios
- Customers with their own defect samples who need a dedicated detection model
- Organizations that can budget for compute investment and training efficiency
- Cloud or on-premise deployment both support training; choose by need
5. Notes
Self-training results depend on sample quality and quantity; it is recommended to proceed under professional guidance. Compute investment should be evaluated against business scale.