Geraldine/shelf-photos-batch1

shelf-photos-batch1

Working dataset for the shelf-monitoring pipeline: bootstrap labeling and fine-tuning data management for Geraldine/rf-detr-nano-bookshelf.

Layout

• —photos/ — 25 of the library's own shelf photos, untouched (no crops, no upscaling)
• —original_images_library/ — 285 high-res images from llabres/library-dataset (MIT), used for continuous fine-tuning (domain shift: real library stacks)
• —dataset/Bookshelf-recognition-2.v1i.coco.zip — COCO export of Roboflow rateshelf/bookshelf-recognition-2 v1 (training data of the current model), cached here
• —annotations/ — auto-label proposals, not ground truth:
• —owlv2_proposals*.coco.json + previews*/ + stats*.json — OWLv2 3-class proposals on photos/ (v1–v3)
• —library-dino-v1-smoke/ — Grounding DINO smoke test (10 library images)
• —library-dino-v1/ — Grounding DINO book proposals on the 285 library images: grounding_dino_proposals.coco.json (COCO 1.0, importable in Label Studio / CVAT), previews/, stats.json
• —gap-analysis-v1/ — gap geometry previews on photos/
• —scripts/ — labeling and run scripts (autolabel_books.py, run_autolabel_library.py, shelf_tiling.py, owlv2_label.py, …)

Pipeline

auto-label proposals → human correction (Label Studio / CVAT) → corrected COCO → adaptive fine-tune from Geraldine/rf-detr-nano-bookshelf (domain images oversampled ×3–5 + recall batch). Never fine-tune on uncorrected proposals.

Note on shelf-end labels

Keep a face-on, full-resolution photo of every range label; those are the crops the call-number OCR stage trains/evaluates on.