Analyze an image or video source for a text-described object, inspect bounded time-sampled rectangular boxes, adjust confidence and display style, or clear the analysis.
Where: Native Video Editor → select a layer with a saved Object Detection effect → Inspector → effect controls.
Key ideas
- Detection boxes are not segmentation masks: The effect draws rectangular detection boxes and can pixelate their regions. It does not trace object edges; use Object Selection when an editable boundary mask is needed.
- Analysis samples a bounded time range: Current frame analyzes one frame. Short range samples a bounded interval around the playhead, at a limited cadence and never beyond the layer’s source interval.
- Detection runs locally: The detector uses the in-browser text-grounding model on the device; it is distinct from paid cloud media generation and does not create a compositor MCP tool call.
Steps
- Select an image or video layer that already has the Object Detection effect. This effect is retained for saved-project compatibility and is not currently offered as a new effect in the Add Effect catalog.
- In the effect controls, enter a short description such as “person”, “car”, or “product”. Detection requires an image/video source and a non-empty query.
- Choose Current frame to analyze the playhead frame, or Short range to sample a bounded range around it. Watch the progress and use Cancel if the analysis is no longer needed.
- Inspect the visible-box count and scrub within the analyzed source interval. Results do not extend beyond the analyzed samples.
- Adjust Confidence to hide weaker candidates. If candidates exist but none are visible, try a more specific query or lower the confidence threshold.
- Adjust box color/width for visibility, or enable Pixelate objects and choose block size when the detected regions should be obscured.
- Choose Clear analysis to remove the stored detections and start over with another query or range.
- For a precise silhouette/matte instead of rectangles, use the separate Object Selection tool and create an editable mask.
Effect controls
- Object description: text query used by local grounding.
- Current frame / Short range: single-frame or bounded temporal analysis actions.
- Confidence: minimum detection score required for a box to render.
- Box color and Box width: visible rectangle styling.
- Pixelate objects and Pixel size: optional mosaic treatment for box regions.
- Cancel and progress are shown while a run is active; Clear analysis removes the sampled results.
Scope and API boundary
- The stored samples are tied to the source asset and source times, so moving/trimming footage does not convert them into a mask or a persistent object track.
- The feature is UI-driven through ObjectDetectionProvider. There is currently no dedicated compositor AI/MCP operation to initiate its analysis; generic project editing should not be described as invoking the detector.
- The Object Detection effect is marked legacy in the effect catalog: existing saved instances can be inspected/used, but the picker does not offer creating new instances.
Tips
- Use a concise noun phrase for the query and analyze a short range first; broad queries often produce many low-confidence boxes.
- Prefer Object Selection or dedicated tracking tools when you need edge precision or a continuous tracked subject.
Limitations and important notes
- Boxes are rectangular and less accurate at object edges than segmentation masks.
- Analysis is limited to image/video layers and the selected current/short source-time interval.
- The legacy effect is not exposed as an MCP-triggerable analysis tool and is not available as a new picker option.
Troubleshooting
Current frame and Short range are disabled.
Select a layer with an image or video source and enter a non-empty object description.
No boxes appear after analysis.
Try a clearer text query or lower Confidence; confirm the playhead remains within the analyzed interval and that the source asset is still the same.