Vision Events

Record, search, and analyze what your deployed computer vision models see in production.

Overview

Vision Events is a Roboflow datastore that can capture important events, like the detection of defect or a count of inventory, alongside visual information from your deployed computer vision model. This gives you a searchable, filterable history of everything your vision system observes in production.

Vision Events replaces Model Monitoring. If you are using Model Monitoring today, see (Legacy) Model Monitoring for the previous documentation.

What is a Vision Event?

A Vision Event is a timestamped record created when a model processes an image. Each event can optionally capture:

• Images: The original images that was processed and associated output images

• Predictions: Object detections, classifications, instance segmentations, or keypoints returned by the model

• Source metadata: Which device, stream, or workflow generated the event

• Custom metadata: Key-value pairs you define for your domain (e.g. line_number, shift, part_number)

Key Concepts

Use Cases

A Use Case groups events that share a common purpose and custom metadata structure. Events in the same Use Case typically send similar metadata fields so you can filter and query them consistently.

For example, a "Defect Detection" Use Case might always include line_id, shift, and part_number even if events come from multiple factories or camera locations. Create separate Use Cases when the metadata structure is fundamentally different (e.g. "PPE Compliance" tracks zone and alert_type instead).