AI Image Tagging¶

Image Tagging describes what is in your video frames using AI, adding the results as searchable keywords. It is the most general of the three AI options: where scene classification picks a single setting for the frame and object detection locates specific objects, image tagging describes the frame as a whole — scenes, objects, settings and activities together. It works from a fixed vocabulary of 141 tags, covering the subjects that make up most video libraries — for example beach, kitchen, crowd, airplane, sunset, construction site.

It runs from the AI Scenes batch tool, alongside scene classification and object detection. Detected tags become orange keywords on the thumbnail and video.

Image tagging can run two ways: as a batch over videos you already have, or automatically as new videos are indexed. Both are covered below.

Download the model (first use)¶

Image tagging uses a model that is downloaded on demand, so it is not part of the installer. The first time you want to use it:

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

The model is stored per user and works without admin rights. Once downloaded, the Image Tagging option becomes available in the AI Scenes tool.

Tag your videos¶

1. Right-click a video (or use the start page) and choose Classify Scenes... to open the AI Scenes tool.
2. Tick Image Tagging. You can combine it with scene classification and object detection, or use it on its own.
3. Set the Tag sensitivity. Higher values give fewer, more confident tags; lower values tag more freely. The default (3.0) is a good starting point, and small changes have a large effect — going from 3.0 down to 2.0 takes a typical frame from around 4 tags to over 40. Move it in steps of 0.25 and re-run a handful of videos to see the result before applying it to a whole catalog.
4. Choose which videos to process and the frame skip, then click Start.

Tags appear as orange keywords. Search for them like any other keyword, and manage them in the Keyword Manager (they show up with the type Tag).

Tag automatically while indexing¶

To tag new videos as they are added rather than running a batch afterwards, open Preferences > AI Classification and tick Image Tagging. The Tag Sensitivity on that same tab is the threshold used for automatic runs, and Frame Skip is shared with scene classification and object detection.

The same on/off switch also appears on the Video Indexer tab as Enable Image Tagging, but the AI Classification tab is where all three are tuned.

The option only has an effect once the model has been downloaded; until then indexing simply skips tagging. Tagging is heavier per frame than scene classification or object detection, so expect indexing to take longer with it switched on — that is the trade for not having to run a batch later.

While processing, each tag is listed with a z-score — a clear subject might score around 5, while a borderline one sits just above your sensitivity value.

How a tag is chosen¶

Understanding this one rule explains most surprising results.

A tag is not added simply because the thing is present in the frame. It is added when the frame scores unusually high for that tag compared with that tag's own normal range. Every one of the 141 tags carries its own baseline, and the z-score you see in the log is how far above that baseline the frame landed. Anything above the Tag sensitivity is kept.

This is why distinctive subjects tag easily and everyday ones may not. A sunset is dramatically unlike the average frame for the tag sunset, so it fires immediately. People, by contrast, appear in a very large share of all video frames, so the tag person has a high baseline — an ordinary shot of somebody standing there is unremarkable for that tag and may not clear the threshold.

It adapts to your footage¶

Those baselines describe what "normal" looks like, and they start out describing general video rather than your video. As you tag, Fast Video Cataloger learns what is normal for your own catalog from the frames it is already processing, and the tags get more accurate as a result. There is nothing to switch on and nothing to wait for — it costs no extra time.

This matters more than it sounds. Baselines built from general video judge your catalog against material nothing like it: whatever is everyday in your own footage is still treated as unusual, so tags can attach to frame after frame for no better reason than being typical of your library. Learning from your catalog removes that. The trade is that on a catalog where every video looks much like the next, less stands out and tagging becomes more sparing than it used to be.

Two practical consequences:

• The first pass over a brand-new catalog has the least to go on. If the results disappoint, re-run tagging on the same videos — the second pass benefits from what the first one learned.
• It is per catalog, so a catalog of interviews and a catalog of drone footage each get judged on their own terms.

That trade-off is deliberate: if tags fired on presence alone, person would end up attached to most of your library, and a keyword that matches almost everything is not useful to search. The threshold is what keeps the tags worth searching for.

Lowering the sensitivity is not a good way to get the common subjects back. The tag count grows far faster than those subjects appear. Measured over a mixed catalog, dropping the sensitivity from 3.0 to 2.0 takes the average frame from under one tag to about four, while person still only reaches a few percent of frames. You get a great deal of noise and not much of what you were after.

To find people, use the tools built for it: object detection includes person among its 80 types and fires on presence rather than distinctiveness, and face recognition identifies who it is. Image tagging is for describing the scene around them.

Can I add my own tags?¶

Not to image tagging — the 141 tags are fixed, and no sensitivity setting will produce a tag that is not on the list. But there are two supported routes for subjects of your own:

• Trained Tags — teach a keyword from around ten example images, right inside the program. No dataset, no training run. This is the right first stop for most custom subjects, from my studio to a particular dog.
• A custom object detection model — the expert path: train a detection model on your own annotated classes and import it. The right answer when you need reliable detection of specific things in quantity — particular machinery, products, equipment, uniforms.

Neither replaces image tagging, though. It covers settings and activities that are not objects at all, such as sunset or construction site. Most catalogs are best served by running them together.

Tips¶

• Tagging looks at the whole frame, so it captures scenes and settings, not just distinct objects — a good complement to object detection.
• If a catalog tags too much or too little overall, adjust the Tag sensitivity and re-run; there is no need to re-download anything. Change it in small steps — the effect is much stronger than the numbers suggest.
• Tagging is heavier per frame than object detection. Use the frame-skip setting to keep large catalogs fast — you rarely need every thumbnail tagged.

Troubleshooting¶

The Image Tagging checkbox is greyed out -- the model isn't downloaded yet. Get it from the AI Models dialog (see above).

Too many or too few tags -- adjust the Tag sensitivity slider and re-run. Higher = fewer, stronger tags.

Common subjects like people or faces aren't tagged -- because they appear in so many frames, the AI needs a stronger-than-usual signal to tag them, so they come back empty at the default sensitivity. This is expected; see How a tag is chosen. Reach for object detection or face recognition rather than lowering the sensitivity, which adds far more noise than it recovers.

A subject I need isn't tagged at any sensitivity -- the vocabulary is a fixed list of 141 tags, so a subject outside that list will never be tagged however low you set the threshold. For subjects specific to your work, teach a trained tag from example images, or train a custom object detection model.

Some clips have far fewer tags than others -- check the frame skip before the sensitivity. Tagging reads your captured thumbnails, and the frame skip decides how many of them it looks at -- at 5, a short clip with few thumbnails comes back thin. Set the frame skip to 1 and re-run those clips.