Object Detection

Input and output format for the default Object Detection demo model.

This AI model, based on YOLOv4-tiny, is trained to detect a diverse set of objects in a given image. See the full list of categories in the COCO dataset.

Input

The input tensor dimension for this model is 320 wide by 320 high.

This model accepts the following extra input information:

• Mask: excludes or includes a specific region of the camera view.

• Sensitivity threshold: NMS sensitivity.

Output

Each detected object is identified using a bounding box represented by a vector: [x1, y1, x2, y2, score, class]. The object bounding boxes contain the following information:

• Position and size coordinates, to be used as a rectangular bounding box.

• A confidence score, which helps retain only the most confident and accurate detections while suppressing weaker or redundant ones.

• Class, a textual string of the main category the detected object belongs to.

The following image shows an example of this model's output:

Example object detection output in the desktop client

Postprocessors

This model's detections can be combined with any of the built-in postprocessors, for example:

• Loitering Detection: detect when an object stays in frame longer than a configured time threshold.

• Left Behind Object Detection: detect objects left behind or removed compared to a reference frame.

• Line Crossing Detection: detect or count objects that cross a defined line, and generate an event.

• Object Counting: count bounding boxes per class and generate a counting event.

See Postprocessors for the full list and configuration details.