shelf-photos-batch1

Working dataset for the shelf-monitoring pipeline behind Geraldine/rf-detr-nano-bookshelf: the images, the auto-generated labels used to build and adapt the model, and the scripts that produced them.

Layout (post-cleanup, 2026-09-25)

Path	What it is	Size	Status
annotations/library-vlm-v1/	⭐ The active labels. 19,671 book boxes on the 285 library images, auto-generated by the frontier VLM GPT-6 Astra (0–1000-normalised grid, cross-checked against independent Grounding-DINO proposals, visually triaged). Contains: vlm_proposals.coco.json (COCO 1.0) · stats.json · 285 annotated previews/ · merged_stats.json · and the merged fine-tune dataset — merged/train_annotations.coco.json (1,772 images / 79,555 boxes: domain ×4 oversample + 800 replay, seed 42) and merged/valid_annotations.coco.json (42 images / 3,116 boxes)	~462 MB	active — unique to this repo
annotations/library-vlm-v1-smoke/	Astra smoke test (1 image, 56 boxes): the record of the coordinate-convention validation (norm1000 confirmed by x-axis cross-check against DINO)	1.6 MB	active — provenance record
original_images_library/	The 285 images the labels point to — required by the COCO files above (the labels are coordinates; without the images they are unusable). From llabres/library-dataset (MIT); mirrored here because upstream only hosts SharePoint links	406 MB	mirror (unstable upstream)
photos/	25 of the library's own shelf photos, untouched (no crops, no upscaling)	2.2 MB	unique — exists nowhere else
dataset/Bookshelf-recognition-2.v1i.coco.zip	COCO export of Roboflow rateshelf/bookshelf-recognition-2 v1 — the stage-1 training data, and the source of the replay batch (--replay 800) used by scripts/finetune_shelf.py against catastrophic forgetting	396 MB	mirror (Roboflow Universe)
annotations/library-dino-v1/	Grounding DINO proposals on the 285 images — COCO + stats only (its 285 full-res previews were removed on 2026-09-25). No longer used as labels; kept as the DINO↔Astra cross-check record cited in the model card	~5 MB	archive
gap-analysis-v1/	Early gap-geometry previews on photos/ (superseded by the live demo Space)	0.5 MB	archive
scripts/	Everything that produced the above: vlm_autolabel_books.py (GPT-6 Astra labelling), autolabel_books.py (Grounding DINO), run_autolabel_library.py, finetune_shelf.py (stage-2 adaptive fine-tune), shelf_tiling.py, owlv2_label.py, derive_gaps_label_retry.py, gap_analysis_roboflow.py	0.1 MB	active

Pipeline status

1. ✅ Stage 1 — fine-tune on the Roboflow domain (bookshelf-recognition-2, human labels), mAP@50 0.974.
2. ✅ Stage 2 — library-domain adaptation on the 285 images with the Astra auto-labels (this repo), pushed as model card version v2-library-vlm-adaptation, mAP@50 0.6755.
3. ⏳ Pending — a hand-checked hold-out (15–20 images corrected in Label Studio / CVAT) to replace the VLM-agreement validation split with a human-ground-truth metric.

⚠️ annotations/* are machine-generated labels, not ground truth. The Astra labels were accepted after numeric cross-check and visual triage; a corrected COCO from the pending human pass would supersede them.

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 (planned, deferred) trains and evaluates on.