Catalog a reference library from a shelf video
Extract visible book titles from a short video and build a reviewable register with frame-level evidence.
How this works
Use this when an office, studio or training team needs a searchable register of physical reference books. A slow shelf video supplies visual evidence; AI proposes titles and authors while retaining the frames that support each identification.
The result is a candidate inventory, not proof of ownership or complete stock. One book appearing across many frames is one observed copy. Similar spines, hidden editions and unreadable labels should remain unresolved instead of being filled with plausible bibliographic details.
Before you start
Capture plan
Record each shelf in order, with steady motion and readable spines; avoid people, private documents and screens in the frame.
Catalog policy
Decide whether to track editions, physical copies, shelf locations and optional ISBNs, and define duplicate rules.
Prepare the inputs
- shelf_video
- Required
Video with capture date and shelf labels
Check resolution and coverage; preserve the original file unchanged.
- existing_register
- Optional / conditional
CSV or workbook
Include stable item IDs and locations when reconciling an existing collection.
- field_schema
- Required
Catalog column definitions
Mark optional fields; unknown ISBNs or editions must remain empty.
Implementation, step by step
Map the video coverage
- Inspect the recording and identify shelf segments, camera reversals and gaps. Create a coverage note before extracting records.
- Sample frames at readable intervals and retain their timestamps. Use available video tools or manually supplied still frames when extraction tooling is unavailable.
Checkpoint: Each observed shelf has a clear evidence range and any missing areas.
Extract visible text
- Read titles, author names and visible edition details from candidate frames. Preserve uncertain readings alongside their image references.
- Do not use a language model’s familiarity with a book to replace unreadable text. A proposed title needs visual support or a reviewer-confirmed match.
Checkpoint: Each candidate contains observed text and its evidence frame.
Resolve identity and duplicates
- Merge repeated observations of the same spine across adjacent frames. Use shelf position and neighboring items to distinguish repeated views from multiple physical copies.
- If permitted, consult a bibliographic source to suggest exact editions, but keep external matches separate until identifying fields agree.
Checkpoint: Candidate items have stable IDs and explicit duplicate decisions.
Build the review queue
- Populate the register with confirmed fields, shelf location and confidence notes expressed as concrete uncertainties.
- Create a short list of items needing a closer photo or manual inspection. Avoid a single confidence percentage that hides whether the title, author or copy count is uncertain.
Checkpoint: The reviewer can resolve each uncertain field without rescanning the whole video.
Reconcile and export
- Compare confirmed items with any existing register. Classify additions, possible moves and not-observed items; not observed does not mean removed.
- Export the catalog and frame index, preserving review status. Apply authorized register changes only after the owner confirms identity and counting rules.
Checkpoint: The result remains traceable to physical observations and review decisions.
Decisions along the way
Two copies appear on different shelves
Keep two observed-copy records unless evidence shows the camera revisited the same location.
A spine is unreadable
Request a closer frame and leave the identity unresolved.
What you will produce
reference-library.csv
CSVA searchable candidate register.
Include: item_id · title · author · edition · shelf · copy_status · frame_ref · review_status
capture-gaps.md
MarkdownA targeted rescan and verification list.
Include: shelf · timestamp · uncertainty · required_photo
Check the result
- Check duplicates against video position and time, not title alone.
- Confirm bibliographic enrichment does not overwrite observed evidence.
- Review all inferred removals and uncertain editions.
When something goes wrong
Fast motion blurs several shelves
Request a focused rescan and mark coverage incomplete.
Frame extraction fails
Use supplied still images and state the reduced coverage.
A worked example
SYNTHETIC EXAMPLE · NOT A REPORTED BUSINESS RESULT
Request
Synthetic: catalog one office shelf.
Inputs
The same blue book appears in five adjacent frames; a second blue copy is visible beside it in the clearest frame.
Expected result
Create two observed-copy rows with the shared title if readable, not five rows; keep the edition unknown if its text cannot be seen.
Run a small pilot
Start small
Pilot one shelf and compare against a manual inventory.
Record the baseline
Measure title errors, false duplicates and minutes per verified item.
Agree success criteria
Require acceptable identity accuracy and evidence for every accepted record.
Stop or revise
Stop bulk import when capture quality cannot support copy counting.
Keep the scope clear
- This adaptation concerns a business reference library; it does not establish ownership, availability or lending rights.
- Do not upload footage containing unrelated confidential material.
Check before using
The library or office owner confirms ambiguous titles and physical copy counts.
Business activities
- Maintain shared company knowledge
- Inspect and maintain physical assets
FROM A RECIPE TO YOUR WORKFLOW
Bring the company context together.
This workflow needs reliable inputs before it needs a better prompt. Moat8 Vault brings working information into a readable archive for the AI tools you use.
For this recipe: Shelf video · Catalog field specification · Existing inventory
Start with one recurring task, agree which sources to use, and review the first result with your team.
Discuss this workflow