Task contract. Keep one complete document page static in Input while Target cumulatively reveals only the official DocLayNet bounding boxes and semantic role labels. The fixed role schedule is a visualization order, not inferred document reading order.
1,200audited Final samples
4,800four-file objects
12Gate C review pairs
READYlifecycle
Current lifecycle status. Gate B v0.3.0 passed both human checkpoints and PR #479 merged. Full audit, canonical Final publication, 12-sample independent reproduction, Gate C Validation publication, and hosted-byte verification all PASS. Gate C Owner and MJXWANG Evaluator PASS are bound to exact head fa3d49c7; promotion PR #525 merged into dev as 1623213e. Task 239 is now READY · MERGED.
Input → target preview
Accepted Gate B continuity preview. Input remains the unchanged page. Target begins clean and cumulatively adds official boxes and role labels in the fixed task-level role order. The current Gate C page below reviews 12 exact-Final pairs.
1. Financial report — table-heavy layout
Task_239_document_layout_role_visualization_000000600 · 17 annotations · 4 roles · 84 frames · 12 FPSINPUT: unchanged complete page
TARGET: cumulative Table, Section-header, List-item, and Text overlays
2. Government tender — header/footer coverage
Task_239_document_layout_role_visualization_000001147 · 31 annotations · 5 roles · 96 frames · 12 FPSINPUT: unchanged complete page
TARGET: cumulative body, header, and footer role overlays
3. Laws and regulations — dense small-box coverage
Task_239_document_layout_role_visualization_000000585 · 42 annotations · 3 roles · 72 frames · 12 FPSINPUT: unchanged complete page
TARGET: cumulative Section-header, List-item, and Page-footer overlays
These six embedded videos preserve accepted Gate B continuity examples. Open the immutable 12-pair exact-Final Gate C review.
Identity and provenance
| Dataset | DocLayNet core 1.0.0 |
|---|---|
| Official source | DocLayNet repository and publisher-hosted DocLayNet_core.zip |
| Published scale | 80,863 pages: 69,375 train, 6,489 validation, and 4,999 test |
| Admitted scope | 1,200 unique annotated train pages selected by deterministic category-aware round-robin |
| Current Raw | s3://tv2v-data/raw/taxonomy/Task_239_document_layout_role_visualization_data-pipeline/ |
| Canonical Final | s3://tv2v-data/final/taxonomy/Task_239_document_layout_role_visualization/; inventory a921c58f…9ea7 |
| License | CDLA-Permissive-1.0 recorded by the publisher |
Measured archive inspection
| Official core archive | 30,012,083,650 bytes; ETag 52242220284e12538949c5b06eb2777c-448; 81,478 ZIP members; HTTP Range support PASS |
|---|---|
| Validation annotation member | COCO/val.json, 59,467,367 bytes, CRC32 f7074b06, SHA-256 eacd2ba2… |
| Frozen train manifest | 1,200 records; SHA-256 25cded58d80bed1d36db4a190e48ee12cab03ea8ec7e381b950b77bc8b587ce5; non-null S3 VersionId |
| Raw inventory | 1,203 objects total; 1,200 PNG pages plus frozen manifest, selection report, and inventory |
| Category census | financial reports 201; government tenders 199; laws/regulations, manuals, patents, and scientific articles 200 each |
Paper, authors, and citation
| Paper | DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis |
|---|---|
| Authors | Birgit Pfitzmann et al. |
| Attribution | Cite the DocLayNet paper and retain the publisher's CDLA-Permissive-1.0 notice with internal and derived artifacts. |
Raw location, acquisition, and contamination
| Source format | 1025×1025 RGB PNG pages plus COCO JSON bounding-box annotations for 11 document roles |
|---|---|
| Canonical Raw | s3://tv2v-data/raw/taxonomy/Task_239_document_layout_role_visualization_data-pipeline/ |
| Fresh acquisition? | Yes; publisher-hosted core archive identity verified before exact page extraction |
| Known modifications | Pages retain complete visual content. Raw stores canonical reduced annotations, exact page and annotation SHA-256 identities, provenance, and deterministic ordinals. |
| Privacy | Pages originate from public documents and can contain names or organizational identifiers. Public review excludes obvious signatures and direct personal contact information; Gate C public-review preflight passed 12/12. |
Native formats and annotation semantics
| Native item | Format | Meaning |
|---|---|---|
| Page | 1025×1025 RGB PNG | Complete document page and task input content |
| Layout annotation | COCO [x,y,width,height] bbox + category ID | Official geometry and one of 11 semantic roles |
| Provenance | Collection, document name, page number, category, image ID | Stable upstream page identity; not an inferred reading order |
DocLayNet core 1.0.0
├── PNG/<page_hash>.png → complete page
└── COCO/train.json
├── images[image_id] → page identity and provenance
└── annotations[image_id] → official bbox + role only
V2V conversion
| Output | Construction |
|---|---|
input_video.mp4 | Static 1024×1024 deterministic area-resize of the complete page for the full clip. |
target_video.mp4 | Same unchanged page, 12-frame clean hold, then cumulative 12-frame role stages and a 24-frame complete hold. |
prompt.txt | Requests official layout-role boxes and labels, explicitly forbidding reading-order inference and unannotated regions. |
metadata.json | Metadata 1.2 with exact versioned Raw source, page/annotation hashes, provenance, role facts, media identity, fingerprint, and producing commit. |
Quality gates
- Underlying page pixels remain unchanged between aligned Input and Target frames except for the intended overlay.
- Every overlay box uses one official bbox and official role; no OCR, reading order, or unannotated region is inferred.
- Coordinates are scaled from 1025×1025 COCO space to 1024×1024 video space and clipped only at visible boundaries.
- All 2,400 Full videos decode completely as H.264/yuv420p fast-start at 1024×1024 and 12 FPS.
- All 1,200 samples pass four-file, Metadata 1.2, unique source/fingerprint, semantic, source-pairing, and media checks.
- Standard resume skips all 1,200 samples with zero failures; independent reproduction matches all 48 selected payloads byte-for-byte.
- Final publication used conditional no-overwrite writes and exact VersionID SHA-256 readback for all 4,800 payload objects.
Pipeline → final four-file contract
publisher core archive → bind HTTP/ZIP identity → exact train PNG + COCO pair → deterministic category-aware selection → versioned Raw page and annotation hashes → 1024×1024 static Input → cumulative fixed-order official-role Target → H.264/yuv420p fast-start pair → prompt.txt + Metadata 1.2 metadata.json → four-file, semantic, privacy, decode, reproducibility, S3, and HTTP checks → audited Full + conditional publication: final/taxonomy/Task_239_document_layout_role_visualization/
Completeness and source verification
| Check | Result |
|---|---|
| Official source/paper/license | Verified |
| Archive and annotation identity | HTTP identity, ZIP census, CRC32, SHA-256, and real-page decode recorded |
| Frozen Raw identity | 1,200 unique train pages; exact manifest, inventory, and VersionIds bound |
| Production Final | 1,200 samples / 4,800 objects / 488,293,204 bytes; Full audit and exact-VersionID readback PASS |
| Gate C review | 12 exact-Final pairs across all six categories, all eight Gate B continuity samples, beginning/end ordinals, seven-role/Footnote, dense-layout, and boundary-clipped-box risks |
| Workflow 3.5 machine checks | PASS; evidence dde29624…3416, 25/25 hosted resources verified |
| Human review | Owner and MJXWANG Evaluator PASS at exact head fa3d49c7; PR #525 merged as 1623213e |