Overcast enhance and resolve

overcast-enhance-and-resolve

Use this skill for the "zoom in… enhance" task: a plate, a face, or on-screen text is too small or noisy to read, and you need to recover it and cite it honestly. Use the broad overcast skill and overcast/reference/verbs.md for exact flags.

Workflow

1. Ingest the raw clip and pin the moment worth resolving:

overcast doctor --json
overcast case init --json
overcast watch ./raw.mp4 --json
overcast note "plate unreadable, want to resolve" --ref <watch-record-id> --at 41-44 --json

2. Enhance that segment. The bundled ffmpeg ops are denoise, normalize, voice-isolate, upscale, stabilize, grayscale; the enhanced file comes back as a media.enhanced record you chain forward:

overcast enhance ./raw.mp4 --ops denoise,upscale,stabilize --json

3. Re-read the enhanced output. --ocr recovers text (a caption/OCR record, no boxes); --detect locates a region and needs a bound detector (bind OWLv2 as the see provider first) — it produces the record with boxes that crop cuts from:

overcast see frame://<enhanced-record-id>@<seconds> --ocr --json                 # -> <ocr-record-id> (text, no boxes)
scripts/visual-db-uv.sh --detect     # once: uv-installs torch+transformers+scipy, prints DETECT_PY
export DETECT_PY="$DETECT_PY"; overcast provider setup apply --preset owl-local --yes --json  # owl-local persists a portable shipped: ref for detect.py + uses $DETECT_PY (the venv python; system python3 lacks the deps)
overcast see frame://<enhanced-record-id>@<seconds> --detect "license plate, text" --json  # -> <detect-record-id> (boxes)

4. Materialize the resolved region as durable cropped evidence — crop the --detect record (the --ocr record has no boxes to crop):

overcast crop <detect-record-id> --all --class "license plate" --pad 0.15 --square --json

5. Record what was recovered with its provenance. State the ops applied and the source record in the finding, keep a before/after note pair, and cite both the raw and enhanced record.id:

overcast note "before: plate illegible at 41-44 on <watch-record-id>" --ref <watch-record-id> --at 41-44 --json
overcast note "after denoise+upscale+stabilize: reads '7ABC123' (2 chars uncertain)" --ref <enhanced-record-id> --json
overcast finding create "plate resolved to '7ABC123' via enhance denoise,upscale,stabilize on <watch-record-id> — 2 chars low-confidence" --ref <detect-record-id> --confidence low --json
overcast brief --export ./enhance-resolve.html --json

Output

The recovered text/object with an explicit confidence, the exact enhancement ops applied, the before/after record.id pair, and the cropped evidence path. Frame whatever you recover as a lead to corroborate, not a settled fact.

Caveats

ffmpeg upscale is interpolation — it cannot invent detail that was never captured. Recovered characters are a lead, not proof; mark them low-confidence and corroborate (a second angle, a second frame, context). For genuine AI restoration bind a model provider (overcast provider setup apply --preset fal --yes, ESRGAN / DeepFilterNet) and re-run see on the restored output — then still corroborate. stabilize and upscale change geometry, so re-derive any box/measurement on the enhanced record, not the raw one.