Trained Tags¶

Trained Tags let you teach Fast Video Cataloger keywords of your own from example images. Pick a handful of thumbnails that show the thing you mean — a person's dog, a particular set, a type of shot — and from then on the AI can find it across your catalog and keyword it for you. There is no dataset to annotate and no training run to wait for: teaching is done by right-clicking the examples, and about ten of them give good results.

Matched frames get purple keywords on the thumbnail and video, searchable like any other keyword (they show up in the Keyword Manager with the type Trained).

A trained tag named like a keyword you already have uses that keyword instead of making a purple one. The keyword keeps its own colour and type and everything it was already on, and Not this and Remove and teach as work on it like on a purple keyword. To get a separate purple keyword, rename the old one in the Keyword Manager before the tag is next run.

Trained tags complement the other AI options rather than replace them: image tagging and scene classification work from fixed vocabularies, and object detection finds 80 common object types. Trained tags are for the subjects those lists do not cover — and unlike a custom object detection model they need no training pipeline, just examples.

Download the model (first use)¶

Trained tags use the same AI model as image tagging, downloaded on demand — if you already downloaded image tagging, this step is only a formality and no large download follows:

1. Open AI Models (from the start page's Maintenance section).
2. Under Trained tags, click Download.

The model is stored per user and works without admin rights. A GPU is recommended, as with the other AI features — matching works on CPU but is much slower.

Teach a tag¶

The fastest way is from the thumbnails you already have:

1. Select one or more thumbnails that show your subject (multi-select works — every selected thumbnail becomes an example).
2. Right-click and choose Teach trained tag, then pick an existing tag or New trained tag....

That is the whole loop: find good examples, teach them. Repeat with a few more thumbnails from different videos and angles — about 10 varied examples give good results, and 5 is a workable minimum.

A thumbnail you teach gets the tag's keyword straight away, and so does its video: you have just said what it shows. The rest of the catalog gets the keyword when the tag is run.

You can also manage everything from the Trained Tags window (open it from the start page's Maintenance section): create tags, add example images from files, the clipboard, or by dragging them onto the example list (useful when your best examples are photos rather than catalog thumbnails), remove bad examples, and disable a tag without deleting it. For a tag that applies to a detected object, a supplied image goes through object detection and the example is cropped to the object; when nothing is detected the whole picture is taught instead, since an image you picked yourself is usually a portrait of the subject anyway. The status line says how many were taught that way.

An image the tag already has an example of is skipped rather than taught twice, whichever way it is added -- the same file picked or dropped again, the same picture arriving under another name or off the clipboard, and the same thumbnail taught again through Teach trained tag or Remove and teach as are all recognised. A different region of a thumbnail you already taught from is a new example and still goes through. The message says how many were skipped.

Each example keeps its own small copy of the region it was taught from, so it still shows in the Trained Tags window after the video is re-indexed or the image file is moved or deleted.

Whole frame or inside a detection¶

When you create a tag you can optionally set what it applies to:

• Leave it empty and the tag matches against the whole frame. Use this for scenes, sets, locations, styles — my studio, drone shot. Whole-frame tags are independent: a frame can match several of them at once.
• Pick a detector label (like dog or car) and the tag matches inside detections of that object instead of the whole frame. Use this to tell apart kinds of the same thing — teach labrador and beagle both applying to dog, and each detected dog is compared against both, with only the best match winning. This needs object detection's model, since it supplies the detections to look inside.

What a tag applies to is fixed when it is created; for a different subject, create a new tag.

When you teach a refining tag from a thumbnail, the example is automatically cropped to the detected object. A thumbnail where the object was not detected, or is too small to use, teaches nothing — the notification tells you how many were skipped and why, so you know the teaching landed. Teaching finds the object itself, so the selected thumbnails do not need to have been through object detection first; the only thing a refining tag needs is the object detection model, and the menu greys the tag out with that reason when it is not downloaded. The new-tag dialog suggests the objects already known to be in the selection.

Test before running¶

In the Trained Tags window, select some videos or thumbnails in the main window and click Test. You see exactly what a run would tag, with a confidence value for each match — and nothing is written to the catalog. This is the quick loop for checking whether a tag is ready: test, add a few more examples where it misses, test again.

Which examples are pulling their weight¶

Each example in the Trained Tags window carries a badge when something about it is worth a look, and a line under the examples tells you how many are flagged. Click an example to read why, and what to do about it; the numbers behind it are in the tile's tooltip.

• Would be missed — measured against the other examples, this one scores below the tag's match threshold, so footage that looks like it would be missed. Sometimes it is a wrong subject or a bad crop: remove it. Often it is the honest one: a tag taught mostly from still photos, with a few examples from video thumbnails, flags the thumbnails — because the thumbnails are what matching actually sees and the stills outvote them. Teach more examples from your own thumbnails, or lower the tag's match threshold a step.
• Rejected by a Not this — it looks more like a frame you rejected than like the rest of the tag, and that frame is shown in the tile's corner. Footage like this example is rejected too. If it is a genuine example, teach more like it so the tag learns the difference, or remove that Not this; if it is not the subject, remove it.
• Near-duplicate — practically the same picture as another example, shown in the corner, such as the frames either side of it in one shot. Keep one; a different angle or scene would add something.
• Whole image — a refining tag taught from a picture in which its object was not detected, so the whole picture stood in. Fine for a portrait, misleading for a wide shot.

Tick Worst first to sort the strip by these. After a Test, each match also shows the example it was nearest to, and the strip counts how many matches each example was nearest to — an example that is never nearest to anything is not what is finding your footage. All of this is advice: a varied set of examples is exactly what a good tag wants, so a low number is a question, not a verdict, and nothing is ever removed for you.

Run it on your catalog¶

• As a batch: the AI Scenes tool has a Trained Tags checkbox, alongside scene classification, object detection and image tagging. Run it over selected, filtered or all videos.
• While indexing: tick Trained Tags under Preferences > AI Classification and new videos are matched as they are added. The setting carries over to catalogs shared through the server, so a server indexes with your trained tags too.

Tags that apply to a detector label look inside object detections, so a run with such tags loads the object detection model even when Object detection itself is not ticked — you only need the model downloaded. When object detection is ticked as well, one detection pass serves both, and the object keywords it stores now remember where in the frame each object was.

Running a tag again after you refined it, with more examples or a Not this, also takes it off the scenes it no longer matches, together with its keyword. Untick Skip thumbnails already classified in AI Scenes for that run, or the scenes that already have results are skipped. Keywords are shared by name, so a scene keyword of the same name that you added by hand is removed from those scenes too.

Correct a wrong match¶

When a trained tag lands on the wrong thumbnail, select the thumbnail, right-click the tag's keyword under the scene keywords in the Keywording window and choose Not this. The same Not this is on the thumbnail's own right-click menu in the video's scene list: under Teach trained tag, each tag opens a small menu with Add as example and Not this. It works there on a thumbnail the tag has not matched too, which then just has no keyword to remove. The match becomes a Not this example of the tag — a picture of what the tag is not — and the keyword is removed. The tag then stops matching that frame, and frames that look more like it than like the tag's own examples, without you having to raise the threshold or guess which example caused it. Not this examples sit after the ordinary examples in the Trained Tags window, marked with a badge, and can be removed there if you change your mind. Teaching the same thumbnail as an example again does the same: it replaces the Not this, as Not this on a thumbnail that was an example replaces the example. Manage trained tags... at the bottom of the Teach trained tag menu opens the window. Use Not this on real wrong matches; a stock photo of "not a dog" teaches nothing useful.

When a trained tag lands on the wrong frame, the correction is itself teaching: right-click the purple keyword on the scene and choose Remove and teach as, then pick the tag it should have been (or a new one). The wrong keyword is removed and the frame becomes an example of the right tag, with that tag's keyword on it — so the mistake makes the tags better.

Both corrections also take the wrong keyword off the video when none of its other scenes has it. A video where another scene still carries the keyword keeps it.

Tips¶

• Vary the examples: different videos, angles and lighting beat ten near-identical frames.
• Matching reads your captured thumbnails, so what was never captured cannot match. The default capture width is 320 — picking the 640 preset (Preferences > Video Indexer, Resize video-frame to be) gives the AI features noticeably more detail to work with. It matters most when the subject is small in the frame, because a tag that applies to a detected object sees only that detection's crop of the thumbnail.
• Each tag has its own threshold in the Trained Tags window. The default is deliberately conservative; if a tag misses real occurrences, first add examples, then try lowering its match threshold a little and re-test.
• Deleting a trained tag cleans up after it: its keyword, its matches and its taught examples are removed from the catalog.

Troubleshooting¶

The Trained Tags checkbox is greyed out -- either the model is not downloaded yet (get it from the AI Models dialog) or no tag has examples yet. Teach a few examples first.

A tag never matches -- add more varied examples and use Test on current selection to watch the confidence values. If they sit just under the threshold, lower that tag's match threshold slightly. For a tag that applies to a detector label, also check that object detection finds the object at all — no detection, nothing to match inside — and that the subject is not tiny in the thumbnail.

A tag that applies to a detector label never matches -- both Test and the indexing and batch passes need the object detection model to find the object first. Download it from Manage AI models if it is missing; Test tells you when that is the reason.

Tags are listed as disabled because of a different model -- those examples were taught with an earlier trained-tags model and cannot be matched by the current one. Delete the affected tags and teach them again; this only happens when the underlying model changes between versions.