AlmazErmilov/corgi-core-annotation-drafts

CORGI · core research corpus

652 source photographs · 120 synthetic originals · 12 appearance variants

Public research collection. Real source photographs, CORGI annotation drafts and synthetic images are separate configurations. DCID and Bologna retain their noncommercial restrictions. Publication does not establish expert review or product training admission.

Collection	Photos	What is labelled	Origin
dcid_lithology	100	Ten lithology classes, ten close-ups each	Producer image labels
bologna_facies	82	Original core photographs and six-class facies masks	Producer pixel masks
full_core	10	Material support on complete Lago scans	CORGI draft masks
default	460	Sandstone/shale image labels and material support	Producer classes + CORGI drafts

Download the single annotation atlas. Every new DCID and Bologna record is shown, followed by all earlier draft annotations.

Complete synthetic atlas

Download the combined PDF, 272 pages, approximately 195 MB.

Section	Contents	Pages
Comparison	Six shared design families, different scenes	3–8
A	All 60 procedural scenes and geometry masks	9–128
B	All 60 generated photographs and annotation drafts	129–248
C	12 source conditioned appearance edits and transferred masks	249–272

Neutral white and gray document backgrounds. Class colors retain their annotation meaning. Section A geometry is exact for its renderer; geological interpretation and real photo accuracy remain unverified. B and C remain annotation drafts. Family comparisons are not matched specimens.

The clickable contents on page 2 links to each collection and design family. The Contents link at the top of each detail page returns to the index. PDF bookmarks also open individual records.

Inspect the actual records

DCID is 512 × 512 close-ups, not full core scans. Original labels are retained, including separate sandstone, siltstone, mudstone and shale classes. No new pixel masks are claimed for these records.

Bologna covers continental to shallow marine facies. It is not a labelled turbidite corpus. Mask colour names come from the original README; RGB-to-class conversion has not been validated. A size mismatch is flagged with pixelAligned=false; it must be reviewed before pixel training. Source splits are provenance only, not a validated independent benchmark. Exact duplicate photos are counted once, with occurrences preserved.

Use each annotation at the right level

flowchart LR
  A[Close-up]:::data --> B[Lithology image label]:::label
  C[Core photograph]:::data --> D[Producer facies mask]:::label
  E[CORGI material proposal]:::draft --> F[Correction and expert review]:::review
  B --> G[Grouped evaluation]:::review
  D --> G
  F --> G
  classDef data fill:#edf1f5,stroke:#586879
  classDef label fill:#edf4ef,stroke:#52705a
  classDef draft fill:#f2edf5,stroke:#79648c
  classDef review fill:#fff4df,stroke:#92773c

from datasets import load_dataset
records = load_dataset(
    "AlmazErmilov/corgi-core-annotation-drafts",
    "dcid_lithology",  # or bologna_facies, full_core, default
    split="development",
)

Every new record retains source, original split, label origin, hash and licence. Bologna includes native photo and mask dimensions. Source images are unchanged. All new records have productTrainingAdmitted=false and expertReviewedByCorgi=false.

Existing drafts and remaining work

The 460 CoreLogging masks remain unreviewed proposals: 73 visual triage flags and 338 model scores below 0.5. Scores are not geological confidence. Lago has 10 material masks, seven layer polygons on one 7A window and two manually traced appearance bands on 10A. Lithology and facies remain unknown for those scans. Full scan details.

This collection does not yet contain 100 expert-labelled full core lithology scans or 100 turbidite examples. Those goals require additional source evidence and geological review; cropping or repeating existing images is not counted as new full scans.

Sources and permissions

See source-specific licences. The collection is not covered by one blanket CC BY licence.

• DCID: 168sir/drill-core-image-dataset, paper, CC BY-NC 4.0. Revision and original archive path are stored per record.
• Bologna: Di Martino, Andrea and Amorosi, Alessandro (2023), Sedimentary Facies Analysis and Segmentation, CC BY-NC-SA 4.0. Source photographs and masks are unchanged; source README is retained.
• Lago Argentino: Zenodo 5815107, CC BY 4.0. CORGI added draft masks and selected appearance annotations.
• CoreLogging: Alzubaidi, Fatimah (2021), Lithology dataset, V1, CC BY 4.0. Original tray imagery: Geological Survey of South Australia. CORGI added draft masks and triage.

Attribution and derivation

Material	What was reused	What CORGI added
CoreLogging / South Australia	Original photo patches and producer image classes	Proposed support masks, triage and editable review records
DCID	Original close-ups and producer lithology classes	A documented subset and shared inspection format, no new pixel truth
Bologna	Original photos and producer facies masks	Provenance, dimension checks and shared inspection format
Lago Argentino	Original core scans	Draft support masks and selected appearance annotations
Norwegian shelf synthetic collections	Geological descriptions as design references	Generated textures/images, procedural scenes and annotation drafts

The real collections above contain source pixels, not only visual references. Their citations and licences remain attached to the originals and derivatives. The synthetic scene guides record the geological design references and generated texture ancestry; these are not photographs of the named Norwegian formations. The 60 generated originals also appear in the annotation comparison with draft labels. Those repeated views are not additional independent photographs.

When using this collection, cite the original dataset(s) used, linked above, and the CORGI repository revision. A CORGI citation does not replace original source attribution. Preserve source licence notices in redistributed subsets, previews and annotation derivatives. See LICENSES.md.

Norwegian shelf synthetic extension

60 procedural scenes at 1024 × 1536 pixels with five material classes. Includes 20 closeups, 20 single cores and 20 three-section trays.

Dataset guide · PDF atlas · Validation

Configuration synthetic_core_ncs, split synthetic_train. The targets come from shared scene geometry, not expert interpretation of real photographs. All 60 scenes share five texture ancestors. They are augmentation candidates, not an independent real core benchmark. No real wells, depths, formation labels or reservoir properties are assigned. No improvement on real data has been measured. Existing source licences and restrictions remain in force; no blanket open licence is assigned to the generated extension.

Generated core image extension

60 independently generated images at 1024 × 1536 pixels. Six prompt families use Norwegian shelf settings as design context.

Dataset guide · PDF review atlas · Manifest · Visual triage

Configuration imagegen_ncs, split unlabelled_synthetic. These 60 images have no lithology, facies or pixel ground truth. Prompt names describe intent, not validated labels. They do not inherit masks from the procedural collection. Some images crop specimen ends despite the prompt. Geological review and pixel annotation remain pending; product training admission is false. No real well, calibrated depth or measured real data benefit is assigned. The other five configurations and their source restrictions are preserved.

Synthetic annotation comparison extension

Arm	Records	Relationship to the earlier collections
Visual annotation drafts	60	Earlier generated originals gain draft annotations
Reference conditioned variants	12	New appearance variants reuse procedural source geometry through draft registration

flowchart LR
    A[60 generated photos]:::data --> B[Visual annotation drafts]:::action
    C[12 procedural scenes]:::data --> D[Appearance edits and mask transfer]:::action
    B --> E[Hierarchy, uncertainty and review]:::review
    D --> E
    classDef data fill:#e2eeed,stroke:#547f7b,color:#173732;
    classDef action fill:#e8edf5,stroke:#657e9e,color:#253a55;
    classDef review fill:#f4ead5,stroke:#a28b55,color:#54451f;

These 72 records are derivatives, not 72 independent new originals. The 60 previously generated images retain their original image bytes. The 12 new reference conditioned variants share ancestors and split families with their procedural sources. The underlying source datasets and their existing configurations remain unchanged.

Configuration synthetic_annotation_comparison, split synthetic_draft. Annotation status distinguishes visual drafts from registered source mask drafts. Neither arm is expert reviewed, benchmark gold or admitted for product training. Lithology names are candidate interpretations or inherited synthetic design intent. Facies hypotheses are not facies ground truth. Native masks, unknown pixels, ignored boundary pixels and full annotation metadata remain available for inspection. Derivatives and their source families must stay together when constructing splits.

Offline comparison gallery · PDF review atlas · Integrity receipt

Publicly accessible research assets. No blanket open licence is assigned. Existing source licences and their use conditions remain in force.