Overcast Visual Target Search

overcast-visual-target-search

Use this skill when the task is to locate a visual target across videos, images, or captured case media. Use the broad overcast skill and overcast/reference/verbs.md for exact flags.

Workflow

For a person with a reference image:

overcast doctor --json
overcast case init --json
overcast face ./clip.mp4 --match ./person.jpg --json   # a match ≥75% auto-suggests a finding
overcast finding list --state triage --json            # triage the auto-suggested lead(s)…
overcast finding accept  --json                    # …accept a real match (or `dismiss `)
overcast crop  --all --class face --json
overcast ask "where does the reference person appear, with timestamps and confidence?" --json
overcast brief --export ./visual-search.md --json      # short by default; --full for the per-match timeline

For an object or open-vocabulary target (--detect needs a bound OWLv2 detector — build it once with scripts/visual-db-uv.sh --detect (it prints DETECT_PY), then export DETECT_PY=… and bind via the preset: overcast provider setup apply --preset owl-local --yes, which persists a portable shipped: ref for detect.py and uses the venv python, NOT system python3 which lacks torch/transformers):

overcast see ./clip.mp4 --detect "red backpack" --json
overcast crop  --all --class "red backpack" --json
overcast ask "list target detections with timestamps, confidence, and crop paths" --json

For logos, landmarks, or near-duplicate visual references:

overcast index create refs --type image-ransac --local --json
overcast index add ./reference-logo.png --to  --json
overcast image match ./clip.mp4 --index  --json   # a RANSAC hit auto-suggests a finding
overcast finding list --state triage --json                  # then accept/dismiss the lead

Output

Return timestamped matches, similarity or confidence where available, source record.id, media.at, and cropped evidence paths created by crop. face --match / image match auto-suggest findings — resolve them with finding list --state triage → accept/dismiss so a run doesn't leave an un-triaged queue; the default brief is short, --full for the per-match timeline. State whether the match came from face --match, see --detect, or local image-ransac matching.

Caveats

Face detections are sampled-frame detections, not unique-person counts. Use face --match for a specific person and include confidence caveats. For exact evidence, use crop to materialize local image records, then synthesize with ask and brief.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: kdr
• Source: kdr/overcast
• License: Apache-2.0
• Homepage: https://overcast.video

Install and usage instructions live in the source repository linked above.