Object Detection¶

Object Detection finds specific objects in your video thumbnails and adds them as searchable keywords. It works from a fixed set of 80 common object types — people, vehicles, animals, furniture, food and everyday items — and can find several of them in the same frame.

It is one of the three options in the AI Scenes batch tool, and it is the one to reach for when you need to find footage by what is in the shot rather than by where it was filmed.

Object detection or image tagging?¶

The two overlap, and it is worth knowing which one answers your question:

• Object detection looks for discrete things and locates them. It will tell you there is a car and a dog in the frame. It fires on presence, controlled by a confidence value.
• Image tagging describes the frame as a whole — including settings and activities that are not objects at all, such as sunset or construction site. It fires on how distinctive the frame is for each tag, which is a different rule entirely.

Running both in the same pass is normal and generally worth it — they cost one trip through your footage together.

Run it¶

Object detection can run two ways.

Automatically while indexing. Open Preferences > AI Classification and tick Object Detection. Every video you add from then on is analysed as it is indexed. See the AI Classification preferences.

As a batch over videos you already have. Right-click a video (or use the start page) and choose Classify Scenes..., tick Object Detection, choose which videos to process, and click Start. This is how you apply detection to a catalog that was indexed before you switched the feature on.

Set the minimum confidence before you start. It controls how certain the model must be before an object keyword is added — raise it if you are seeing objects that are not there, lower it if obvious objects are being missed. The frame skip controls how many thumbnails are processed; 1 means every thumbnail.

Where the results go¶

Detected objects become green keywords on the thumbnail and the video, each carrying the confidence value it was detected with. They search, filter and sort like any other keyword, and you can manage them in the Keyword Manager.

Choosing a model¶

Two built-in models are available, and you can switch between them in the AI Models dialog (from the start page's Maintenance section):

Click Use on a model's row to make it the active one. Both recognise the same 80 object types — the difference is accuracy against speed, not vocabulary.

If the 80 built-in types do not cover what your videos contain, you can train and import your own model; see Custom Object Detection Models.

Tips¶

• Detection reads your captured thumbnails, not the video file. An object that never appears in a captured thumbnail cannot be detected, so a wider thumbnail interval means fewer detections regardless of the settings here.
• Use Skip thumbnails already classified when re-running over a catalog, so a second pass only covers what the first one missed.
• Detection is lighter per frame than image tagging, so a smaller frame skip is more affordable here than it is for tagging.

Troubleshooting¶

Nothing is detected at all -- check that the model is present in the AI Models dialog. If the batch tool reports a missing model, download it there.

Too many wrong objects -- raise the minimum confidence. Frames that are blurred, very dark, or heavily motion-smeared produce the most false positives.

Obvious objects are missed -- lower the minimum confidence, and try the YOLOX large model if you are running the small one. Small objects in wide shots are the usual hard case.

The object I need is not one of the 80 types -- no confidence setting will help; the vocabulary is fixed. Train a custom object detection model on your own classes instead.