How to Deploy a Custom AI Model on DJI Matrice 4: A Production Checklist

Step 1: specify the detection task before collecting data

Write the class definition so two annotators would label the same object. “Damage” is usually too broad; “missing red safety cap visible from the inspection route” is testable. Define minimum object size, sensor, expected distance, day or night conditions and the action that follows a detection.

Choose evaluation metrics around consequences. If missing a person or a fire cue is costly, recall may dominate. If each alert stops an automated patrol, false positives impose operational cost. Record the target balance before adjusting thresholds.

• Object class and explicit positive/negative examples.
• RGB or thermal input and intended camera settings.
• Altitude, distance, gimbal angles, zoom and speed.
• Acceptable misses, false alerts and response latency.
• Who reviews the output and what they do next.

Step 2: build representative training and test sets

Collect data from the planned aircraft and route when possible. Include the seasons, backgrounds, lighting, thermal conditions, target sizes and occlusions that the model will encounter. Negative images are essential: they teach the model which confusing patterns are not the target.

Split data by flight, site or time rather than randomly separating nearly identical frames. Otherwise, adjacent video frames can appear in both training and test sets and produce an unrealistically optimistic result. Freeze a test set that is never used to choose hyperparameters or thresholds.

1. Create an annotation guide and label consistently.
2. Remove or control near-duplicate frames.
3. Separate training, validation and test data by mission or location.
4. Track dataset versions and licensing or privacy constraints.
5. Document known gaps before training.

Step 3: train for the target platform, then quantize

DJI’s workflow requires a compatible model and optimization process. Quantization reduces numerical precision and adapts the network for efficient device execution. This can change predictions, especially for small or low-contrast objects, so desktop validation does not carry over automatically.

Keep the original weights, training configuration, dataset version and evaluation report. Upload through the approved DJI developer pathway, use representative example images where requested and identify the resulting device-ready artifact with a new version.

Step 4: validate on the aircraft and in the operator interface

Bench testing confirms that the model loads; flight testing confirms that it solves the job. Use a controlled scenario with ground truth and execute the production route. Measure frame rate, end-to-end latency, precision, recall and duplicated events. Check thermal and RGB feeds separately if both are used.

Observe what the pilot sees. Bounding boxes should not obscure critical flight information. Alerts need a class, time and evidence. Confirm behavior when the target leaves the frame, the link degrades, the mission pauses and the model is changed.

Step 5: control deployment, monitoring and retraining

A production model needs an owner, approved devices, a version record and a rollback process. Log model version with detections. Sample rejected and missed cases. Watch for data drift when seasons, assets, camera firmware or routes change.

Spectro AI develops DJI on-board algorithms for the Matrice 4 series and can train models for defined enterprise targets. If the desired model or workflow exceeds aircraft constraints, the same use case can be evaluated for local inference on Brain-Box.

• Bind releases to approved device identifiers and operating configurations.
• Preserve a signed-off baseline and a tested rollback.
• Review false positives, false negatives and uncertain cases.
• Retrain only with controlled labels and re-run the full acceptance suite.
• Do not silently change thresholds or models in an operational workflow.

Frequently asked questions