Object Detect
Detect and classify objects in images with YOLO models.
What it does
Object Detect runs a YOLO model over an image, draws a box around each object found and reports what it takes each object to be. Three quality levels trade speed for accuracy: Fast keeps up with a live camera feed, Accurate takes longer per frame but catches more objects and misclassifies less.
Alongside the visual boxes it reports how many objects were found, their class labels, and a confidence per object, so other nodes can be driven from the count or the labels without touching the image.
When to use it
Counting, locating or classifying objects in a camera feed or image
An effect triggered only when a particular kind of object appears
An interactive installation needing real-time detection that keeps pace with a live input
Per-object data (position, label, confidence) driving other nodes
| Parameter | Type | Default |
|---|---|---|
image | OneOf([FieldOf(Raster)]) | — |
quality | String | "fast" |
confidence_threshold | Number | 0.500 |
nms_threshold | Number | 0.450 |
enabled | Boolean | true |
line_width | Number | 2 |
color | Color | rgba(0.00, 1.00, 0.00, 1.00) |
opacity | Number | 1 |
Gotchas
Confidence Threshold filters out anything the model is unsure about. Raise it against false positives, lower it when real objects are missed.
NMS Threshold controls how aggressively overlapping boxes are merged. Lower it when the same object shows up with two or three stacked boxes; raise it when boxes on separate-but-close objects disappear.
Line Width and Color only style the bounding-box preview. They do not affect detection, and the labels and confidences outputs are unchanged by them.
Toggle Enabled off to pause detection without disconnecting the node, for before-and-after comparison in a live setup.
Worked example
Wire a camera or image into Object Detect.
Leave Quality on Fast for a live feed, or switch to Accurate for a still image where speed doesn't matter.
Raise Confidence Threshold if detections are turning up that you don't trust.
Use the boxes output as a visual overlay, or read count / labels / confidences to drive other nodes, for example triggering an effect only when count is greater than zero.
See also
Vision Analyze · Vision Segment · Vision Tracker · Rasters (images)