dtourolle/scene-actor-extraction
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

0 Releases 1 Tag

RSS Feed
  • v1.0.0 ff3b8ebf1d
    Compare

    v1.0.0 — learned scene-boundary flood-fill pipeline
    Unit tests / Build and run the GPU-free suite (push) Failing after 2s
    Details
    Traceability Validation / Check requirement traces (push) Failing after 9s
    Details

    dtourolle released this 2026-08-11 18:57:13 +00:00 | 2 commits to master since this release

    First tagged release. opencv5 detect/track/match/scene pipeline on KPN++, with
    Jellyfin-integrated gallery and X-Ray-validated tuning.

    Highlights:

    • Flood-fill presence on a learned XGBoost scene-boundary detector: per-second
      X-Ray presence F1 62.6% -> 74.9% (leave-one-out, all nine films improved).
    • Boundary detection at the shipped ±20s tolerance: 44.1% macro leave-one-out
      vs 29.8% grayscale (72.9% on the training films).
    • Ten-knob differential-evolution tuning of the matcher/tracker config.
    • One-clock tracker/registry fix (evidence-watermark), no dropped votes.
    • Scene detector built into scene_analyze via CMake (XGBoost C API + FFTW),
      verified end to end and through the Jellyfin work-queue worker.
    Downloads