Frame-Accurate Scene-Cut Detection

MasterSelects detects hard shot boundaries from every decoded source video frame. The detector records the presentation timestamp of the first frame of the new shot, so results remain frame-accurate for variable-frame-rate and 50/60 fps footage even though the JPEG proxy itself is capped at 30 fps.

Workflow

New proxy generation runs scene-cut detection in the same sequential VideoDecoder pass, before frames are deduplicated to the proxy frame rate.
Existing proxies are left ready and untouched. Regenerate > Scene Cuts in either the Media or Timeline clip context menu starts a decode-only source scan.
Results are stored on the source MediaFile and persisted in the project. The cache contains detector/schema versions, analysis dimensions, decoded and expected frame counts, and a source size/last-modified fingerprint.
Replaced, changed, incomplete, or old-version sources are not accepted as a current frame-accurate cache.

Detector

Every decoded source frame is resized to 160x90. The detector combines:

per-pixel Y/Cb/Cr difference;
the percentage of materially changed pixels;
global Y/Cb/Cr histogram distance;
edge-change ratio with one-pixel tolerance;
a small translation search that compensates camera pans up to eight analysis pixels;
an adaptive threshold based on the preceding two seconds.

Shared-background shot/reverse-shot edits can pass through a combination of content, histogram, edge, and motion-compensated evidence, while the rolling baseline keeps ordinary subject motion below the local cut threshold. Detector version changes invalidate older cut caches and apply improved scoring during rescan.

A candidate is delayed by one frame. This lets the detector suppress an isolated flash that immediately returns to the preceding image. An unconfirmed candidate on the final source frame is not emitted.

The first frame after a confirmed boundary supplies both the exact source timestamp and decoded presentation-frame number.

Performance and platform behavior

When Worker and OffscreenCanvas are available and the Linux safety policy does not apply, source frames are cloned into a dedicated analysis worker. The worker owns its OffscreenCanvas, performs the 160x90 readback and scoring, and applies bounded backpressure so decoded VideoFrame handles cannot accumulate without limit. Other environments fall back to the main-thread software analyzer.

Linux follows the shared Mesa safety policy and uses a main-thread software 2D canvas with willReadFrequently: true. The canvas is fixed at 160x90 and never scales with source resolution or timeline width.

Scene-cut progress is based on decoded source frames, not written proxy JPEGs. Proxy readiness and scene-cut analysis state remain independent.

Timeline display

Detected cuts are drawn as thin dashed black vertical markers on video clips. Their stroke width and dash size shrink as markers become denser while zooming out, so a cut-heavy source does not cover its thumbnails. Marker positions are mapped from the stored source timestamps into each clip’s current in/out range, including live trim previews and reversed playback. Cuts outside the visible source range and linked audio clips are not marked. The same marker painter is used by the worker canvas and the Linux main-thread fallback. The clip Properties Analysis action bar includes a Cuts control that shows the source-wide cut count or decode progress and changes between Analyze, Reanalyze, Retry, and Cancel as its state changes. It sits alongside Focus & motion, Faces, Transcript, Audio intelligence, and AI Scenes. Analyze all coordinates visual analysis, scene cuts, transcript, and audio intelligence; source-decoding jobs share a resource lock, while audio and provider work can run independently. AI Scenes remains an individual action.

Scope

This detector targets frame-to-frame hard-boundary signals. Results are exposed as a persisted cut index, a found-cut count in the context menus, and clip markers. Automatic timeline splitting remains a separate editing action.