Skip to main content

Detection Targets and Regions

Detection target configuration is the step where you define which objects the Vision model should detect in the video. This is a critical setting for efficiently utilizing EVA’s system resources.

Additionally, detailed configurations related to detection targets—such as regions, sensitivity, and reference images—can be intuitively managed through the “Settings & Details” modal accessible from the camera streaming view.



Basics of Detection Target Configuration​

When you first create a detection scenario, objects mentioned in the scenario are automatically extracted and set as detection targets. You can add or modify this list at any time.

Through the Settings & Details option in the camera streaming view, you can also configure:

  • Region of Interest (ROI)
  • Object detection sensitivity
  • Reference images for detection

This allows you to obtain different detection results from the same camera depending on the area of interest and use case.

Model Selection and Sensitivity Adjustment​

EVA provides multiple Vision models, allowing you to select the one that best detects your configured targets, improving analysis quality.

You can also fine-tune object detection sensitivity:

  • Higher sensitivity → Detects only more certain objects (reduces false positives)
  • Lower sensitivity → Detects a wider range of objects (reduces missed detections)

It is important to balance false positives and missed detections based on your operational environment.

💡 VLM is used for contextual reasoning and has higher computational cost, while Vision models are relatively lightweight and suitable for continuous object detection. Therefore, structuring the system so that the Vision model filters detection targets first, and the VLM analyzes only detected cases, significantly improves resource efficiency.



Using Reference Image-Based Detection​

EVA’s Vision models are based on foundation models with strong zero-shot object detection capabilities, enabling high performance on general objects.
However, domain-specific objects (e.g., equipment with unique logos, custom-shaped parts) may not be detected accurately.

In such cases, EVA provides reference image-based detection:

  1. Register Reference Images Add reference images of domain-specific or target objects.

  2. Perform Detection The Vision model uses these images to detect similar objects without additional training.

This feature can be configured directly from the camera streaming view → Settings & Details modal.