FrameQuery 1.2.1: Search Footage by Bib Number

FrameQuery 1.2.1 is out. It reads competitor numbers straight off your footage: race bibs, jersey numbers, the paper number pinned to a dancer's back. Search bib:100 and you get every clip where number 100 appears, dropped at the moment it's on screen. AI processing has moved onto the GPU on both Windows and Apple Silicon. A batch of search filter bugs is gone. Free update for everyone.

What's new in 1.2.1

Bib and jersey number search. Type bib:100 or number:100 and results seek straight to the read.
A one-time library scan. Older footage gains number search from a background pass. No processing minutes used.
AI on the GPU. DirectML on Windows, Core ML on Apple Silicon, automatic CPU fallback.
Filters that stick. Quotes, -exclusions, codec: and tags now apply to AI results too.
A "Reading Bib Numbers" step on the processing card, shown when the feature is on.
The sampled frame cache is visible. It has its own segment in the storage breakdown and a clear button in Data Management.
New dashboard charts. Shot mix, activity by source, weekday rhythm.

Find every clip a competitor appears in

If you film events, you know the email. A client wants everything with couple 100, or runner 4183, and until now the only honest answer was an evening of scrubbing. Now it's a search.

Flip on Competitor Numbers in Settings → AI Models. FrameQuery reads printed numbers from frames it already samples while indexing. Black digits on a white bib is the case it was built for; printed jersey numbers get picked up too, though knit fabric and outline fonts make those less certain. Each read is saved with its timestamp and a confidence score. Search bib:100 and the result card shows the very frame the number was read from. Click it and the player opens right there.

Recognition only considers digits. Generic OCR will happily read an S as a 5 or a B as an 8. Ours can't, because letters aren't on the menu.

We tested on 52 minutes of ballroom competition tape. The scanner pulled out 49 distinct numbers, and the couples actually competing were each read dozens of times at full confidence. One clean look at a bib is all it needs, and at an event, every rotation is another look.

Footage indexed before 1.2.1 needs one pass: Settings → AI Models → Scan library. It runs in the background, resumes if you quit, and doesn't touch your processing minutes. Big libraries take a while because some files have to be decoded again.

All of this runs on your machine, same as faces and voices. Nothing is uploaded.

There's a full walkthrough in the new guide to finding footage by number.

Your GPU now does the heavy lifting

FrameQuery's models used to run entirely on CPU. As of 1.2.1 the heavy ones run on your GPU instead: DirectML on Windows, which covers NVIDIA, AMD and Intel integrated chips alike (anything DirectX 12), and Core ML on Apple Silicon.

Visual search indexing went from 113 ms per frame to 17 ms on our Windows test machine. The part you'll feel day to day is quieter, though: indexing a big library no longer pins every CPU core. The app stays responsive while it works, and so does the rest of your computer.

Not everything moved. We measured each model and some of the small ones are quicker on CPU, so they stayed there. No usable GPU, or a driver acting up? Everything falls back to CPU on its own and behaves like it always did. More on how that works on each platform in GPU acceleration on every platform.

Search filters that always meant what they said

Search pulls results from four places: a keyword engine plus three AI rankings. Filters were only reliably enforced on the first one. So -truck could still return trucks, quoting a phrase sometimes made results worse instead of narrower, and codec:prores next to a text query did nothing at all. It had been that way for a while.

1.2.1 fixes it properly. Filters are resolved once and every result has to pass them, no matter which engine found it. Quotes, exclusions, tags, file metadata, dates, durations. The full operator set is documented in advanced video search operators.

What's fixed

Excluded terms and quoted phrases no longer leak through AI results
codec:, res:, camera: and similar filters work alongside a text query
The content safety panel on a video page showed incidents from other videos; it now shows only that video's
Library scans that read source files again use hardware video decoding where the codec allows
Videos with a missing frame rate recover it by probing the file instead of silently skipping frames
The same number read twice in one frame no longer counts double

How to update

FrameQuery checks for updates on launch. Accept the prompt, or grab the installer from the download page. Libraries need no migration. Bib search is off by default and one toggle away in Settings → AI Models.

Do I need to reindex for number search?

New footage is scanned automatically once the toggle is on. Footage from before 1.2.1 needs the one-time Scan library pass. It runs in the background, resumes if interrupted, and is free.