How to find b-roll by describing the shot

The short answer

Use visual search: a model compares a text description with the frames of your footage and ranks the shots that look most like it. That lets you find b-roll by what's in the picture rather than by file names, folders or spoken words. For shots with text in them, such as signs or screens, on-screen text search catches what visual search misses.

The traditional way: logging

Editors have always solved this by logging: watching every clip and adding keywords, markers or bins. Final Cut Pro keywords, Premiere bins and a good naming scheme all help, and on a well-run project they're still worth doing.

The problem is time. Logging hours of footage takes hours, it's rarely done for old projects, and the keywords only cover what someone thought to write down. When you need "a hand turning a coffee cup" from a shoot three years ago, the log probably doesn't say that.

How describing a shot works

Visual search models are trained on huge numbers of images paired with captions, so they learn to place a picture and a sentence describing it close together. Your footage is sampled into keyframes, each frame is converted into a vector, and your query is converted the same way. The closest matches come back first.

It's good at scenes, objects, settings, lighting and composition. It's weaker at fine detail, specific people you haven't identified, and abstract ideas. Short, concrete descriptions work best. Describe what a camera would literally see, not the story you want the shot to tell.

• Good: drone shot over a forest road, close-up of hands typing, crowd at night with phone lights.
• Harder: a shot that feels hopeful, the take where the lighting was better.
• Combine signals when you can: a scene description plus a word that was on screen, or plus a color or mood.

How to do it in MediaFind

MediaFind indexes pictures, speech and on-screen text together, on your computer.

1. Add your footage folders. MediaFind extracts keyframes and builds a visual index along with transcripts and on-screen text.
2. Describe the shot in plain words, for example aerial of a coastline at sunset.
3. Refine with other search modes: color and mood, actions, on-screen text, or reverse image search with a still you already like.
4. Use related clips on a good result to find similar shots from the same shoot or others.
5. Click to preview at the exact moment, then drop clips on the Create timeline or, with Pro, export your selects as FCPXML or EDL.

Tips

Visual search looks at sampled frames, so a very short shot inside a long clip can occasionally be missed; searching a second, differently worded description often catches it. Results are ranked, not filtered, so scan past the first few before deciding a shot doesn't exist. Photos (JPG, PNG, HEIC and more) are indexed alongside video, so stills and footage come back in the same results. For how the visual index works under the hood, see searching b-roll and semantic visual search.

Frequently asked questions

Can I find footage that has no dialogue?

Yes. Visual search matches your description against the frames themselves, so silent b-roll is just as searchable as interviews.

Do I still need to log my footage?

Logging still helps on active projects, but visual search means unlogged and archived footage is no longer invisible.

Can I search with an example image instead of words?

Yes. MediaFind supports reverse image search, so you can use a still or a frame to find similar shots.

Does my footage get uploaded to analyze it?

No. MediaFind computes keyframes and the visual index on your own computer.