Next Steps

You’ve created a Box Detect function, imported images, drawn bounding boxes in the Train tab, and invoked the function. From here, the work is mostly about scaling up annotation, integrating into your application, and using per-object confidence well.

Use per-object confidence

Every detected object has its own confidence score. Most production detection integrations filter detections by a confidence threshold before acting on them — the right number depends on the cost of acting on false positives vs. missing true positives.

Related: Tuning Confidence Thresholds — written for classification but the general framework applies to per-object detection confidence too.

Plan for production

• Image size and format. Most detection models work best on images within a sensible size range. Very large images may be downscaled; very small ones may lose detail. Test with images representative of production.
• Variable-length responses. Detection responses contain a variable number of predictions. Handle empty arrays, single detections, and dense scenes alike.
• Latency. Box Detect inference is typically more compute-intensive than classification. Plan timeouts accordingly.

Related: Production Integration Patterns.

Go deeper on the platform

• Developer Platform — architecture, principles, and reference for building production integrations.
• API Reference — complete REST API documentation.

Or try another function type

• Create Your Own Classification Function — assign each input to one of a set of labels.
• Create Your Own Search Function — query a reference corpus by meaning.