NLP-based chapter boundary detection with multi-weighted editorial scoring. Every chapter in the footage library gets a composite score across four dimensions, enabling automated rough cut assembly from highest-scoring content.
Each chapter receives a composite editorial score calculated from four weighted dimensions. These weights are tuned for documentary/interview-style content — the primary footage type in the current library.
Dialogue density is weighted highest because the source material is a read-through performance where spoken content is the primary editorial signal. Pacing captures rhythm and beat structure. Content and technical scores evaluate semantic richness and production quality respectively.
The right-skewed distribution is expected — most footage is competent but few chapters hit all four dimensions simultaneously. The top-scored chapters (7.0+) are the best candidates for automated rough cut assembly.
Each chapter gets an auto-extracted thumbnail at its temporal midpoint via FFmpeg. These 27 PNG thumbnails serve as the visual index for the catalogue browser and chapter navigator tools.
| Parameter | Value |
|---|---|
| Extraction Method | FFmpeg midpoint seek |
| Format | PNG (lossless) |
| Count | 27 thumbnails (1 per chapter) |
| Used By | Catalogue browser, Chapter navigator, Library search |
Each thumbnail auto-extracted at chapter midpoint via FFmpeg. Scores shown are composite editorial scores across four dimensions.